Claudia Marzi, James P. Blevins, Geert Booij and Vito Pirrelli Inflection at the morphology-syntax interface
Abstract: What is inflection? Is it part of language morphology, syntax or both? What are the basic units of inflection and how do speakers acquire and process them? How do they vary across languages? Are some inflection systems some- what more complex than others, and does inflectional complexity affect the way speakers process words? This chapter addresses these and other related issues from an interdisciplinary perspective. Our main goal is to map out the place of inflection in our current understanding of the grammar architecture. In doing that, we will embark on an interdisciplinary tour, which will touch upon theoreti- cal, psychological, typological, historical and computational issues in morphol- ogy, with a view to looking for points of methodological and substantial convergence from a rather heterogeneous array of scientific approaches and the- oretical perspectives. The main upshot is that we can learn more from this than just an additive medley of domain-specific results. In the end, a cross-domain survey can help us look at traditional issues in a surprisingly novel light.
Keywords: Inflection, paradigmatic relations, word processing, word learning, inflectional complexity, family size, entropy
1 The problem of inflection
Inflection is the morphological marking of morphosyntactic and morphosemantic information like case, number, person, tense and aspect (among others) on words. For instance, a word may be specified as singular for the grammatical cat- egory of number, i.e. it has a certain value for the feature ‘number’. The feature ‘number’ has two values in English: singular and plural. The choice of specific values for such features may depend on syntactic context or semantic context.
Claudia Marzi, Vito Pirrelli, Italian National Research Council (CNR), Institute for Computational Linguistics, Pisa, Italy James P. Blevins, George Mason University, College of Humanities and Social Sciences, Fairfax, USA Geert Booij, Leiden University, Leiden University Centre of Linguistics, Leiden, Netherlands
Open Access. © 2020 Claudia Marzi et al., published by De Gruyter. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. https://doi.org/10.1515/9783110440577-007 Inflection at the morphology-syntax interface 229
Morphosyntactic features are inflectional features that play a role in syntax. That is, they play an essential role in the interface between morphology and syntax. For instance, the syntax of languages often requires that words in spe- cific syntactic contexts agree with respect to the value for certain features of other, syntactically related words. An example is subject-verb agreement in English: the finite form of a verb has to agree in the values for person and num- ber with those of the subject. Another well-known type of agreement is gender agreement: in many languages determiners and modifying adjectives have to agree in gender with their head noun. Besides agreement, there is a second type of syntactically driven feature value selection, traditionally referred to as government: a word or syntactic construction is said to govern the choice of a feature value for another word. For instance, in many languages nouns have to be marked for case, depending on the syntactic or semantic role of that noun (subject, object, agent, etc.). The grammatical or seman- tic role of an NP then governs the case marking of its head noun. Morphosemantic features are features that are not required by a syntactic con- text, and their choice is primarily motivated semantically. For example, all finite forms of English have a tense property such as present or past. Their choice is not governed by syntactic context, but by what content the speaker wants to convey. Yet, the choice is obligatory, as a specific tense property has to be chosen. In this respect, inflection differs from derivation, which is not obligatory. However, con- text may play a role in the choice of morphosemantic features as well. For in- stance, in a sentence such as Yesterday I went to the movies the past tense form went is normally required because of the presence of the adverb yesterday. Inflection may be divided into two subtypes: inherent inflection and contex- tual inflection (Booij 1993, 1996; Kibort 2010). Inherent inflection is primarily determined by what the speaker wants to express, and is therefore a matter of choice. The speaker determines, for example, the choice between present tense and past tense of verb forms, and the choice of number (singular or plural) for nouns. Contextual inflection is the kind of inflection that is required by syntac- tic context. This is the case for the choice of person and number values for finite verbs in English, which is a matter of agreement. Hence, in English the feature ‘number’ is inherent for nouns, but contextual for verbs. Case marking on nouns may function as contextual inflection in a language like German, where subject nouns and object nouns have to be marked as having nominative and accusative case respectively. That is, this instance of case marking is required by syntax. These are the so-called structural cases, and stand in contrast with semantic case marking, which is a case of inherent inflection. For instance, in Latin we can express the instrumental use of a knife by means of marking the noun with ablative case: cultr-o ‘with a knife’. This is a semantically governed 230 Claudia Marzi et al. case, and hence a matter of inherent inflection. Adjectives in German agree in case marking with their head nouns, and, therefore, case marking is contextual. A survey of the different types of inflectional features is given in Kibort (2010). An example that illustrates what can be expressed by inflection is the fol- lowing sentence of the language Maale, a North Omotic language spoken in South Ethiopia (Amha 2001: 72):
(1) bayí-ské-nn-ó tá zag-é-ne cow-INDF-F-ABS 1SG.NOM see-PRF-AFF.DECL ‘I saw a cow (which I did not know before)’
The word for cow has morphological markers for indefiniteness, feminine gender, and absolutive case, and the verb is marked for aspect (Perfect) and for the sen- tence being affirmative (AFF) and declarative (DECL) in meaning. The ending -ne is the cumulative exponence of the two sentence modalities Affirmative and Declarative. The pronoun for ‘I’ is the nominative form, but there is no separate case marker. This example illustrates some of the formal complications in the ex- pression of inflectional categories, and in particular that there may be no one-to- one mapping of form and meaning in inflection. Inflection may thus considerably increase the formal complexity (discussed in Section 8) of a language system. A third type of traditional inflectional features are purely morphological features such as inflectional classes for nouns and verbs. In Latin, case marking on nouns is performed in five different ways, and hence it is conventional to distinguish five different inflectional classes (declensions) for nouns. Individual nouns are then marked for the inflectional class they belong to by means of a feature. These features are purely morphological because they tend to have no role in syntax or semantics. Similarly, Latin is associated with a number of inflectional classes for verbs, the conjugations. The patterns de- scribed in terms of inflectional classes add substantially to the complexity of a language, and raise the question how children acquire such morphological systems. In many languages the gender of nouns is marked on related words. For instance, in Dutch nouns have either common or neuter gender, and this mani- fests itself in agreement phenomena: determiner and adjective have to agree in gender and number with the head noun. However, the nouns themselves do not carry a morphological marker for gender. Thus, one can only discover the gender of Dutch nouns indirectly, by looking at agreement data. This is another challenge for the language learner. Two interacting but logically distinct issues lie at the core of any linguistic or psycholinguistic account of inflection: Inflection at the morphology-syntax interface 231
A. the issue of what syntactic contexts require morphosyntactic and/or mor- pho-semantic word marking and for what lexical/grammatical units; B. the issue of how morphosyntactic and morpho-semantic information is overtly realized on lexical/grammatical units.
In this chapter we mainly focus on the second issue, the ways in which morpho- syntactic and morphosemantic information is morphologically marked. It must be observed that there is no one-to-one relationship between inflectional features and units of form (‘morphs’) in inflected words, as we will see in Section 2: one morph may express more than one inflectional property (cumulative exponence), and one inflectional property may be expressed by more than one feature (extended expo- nence). Moreover, the same inflectional property may be expressed in a number of different ways. The patterns of interdependent choices are expressed by inflec- tional classes. A simple example from English is that the past tense forms of verbs may be formed either by means of suffixation of the stem with -ed,orbymeansof various types of apophony (Ablaut), i.e. vowel change in the stem, usually with consequences for the form of participles. It does not make any difference for the role of the feature value ‘Past’ in syntax and semantics by which formal means it is expressed. This issue is broached in more detail in Section 2. Given the lack of a one-to-one mapping of form and meaning in the domain of inflection, it is useful to introduce paradigms, systematically structured sets of inflectional forms of words, in order to make the right generalizations and the proper computations (discussed in Section 5). The nature and structure of inflectional paradigms are discussed in Section 3. The inflectional paradigms of words may also contain word combinations. For instance, in Germanic and Romance languages various inflectional forms can be treated as consisting of an auxiliary and a non-finite form of a verb. This is referred to as periphrasis (Section 4). After this brief sketch of the nature of inflectional systems, we will discuss in more detail the way inflection is acquired, how machines can learn it, how it can be modeled computationally and how morphological complexity can be computed (Sections 5–8). Section 9 will summarize our findings.
2 Inflectional syntagmatics
The inflectional features of a morphological system determine observable pat- terns of variation in shape and distribution. In turn, observable patterns of vari- ation cue features. Yet the relations between features and patterns of variation 232 Claudia Marzi et al. are often intricate, typically involving complex interactions between syntag- matic arrangements and paradigmatic classes.
2.1 Morphemes and inflection
Post-Bloomfieldian models seek to establish a tight connection between inflec- tional features and morphotactic units by bundling features and forms into in- flectional ‘morphemes’. In the immediate constituent analyses developed by Bloomfield’s successors, inflectional formatives were included among the ter- minal elements of a syntactic representation, along with bound stems and free forms. The idea of treating inflectional sub-word units as syntactic elements was taken over by generative accounts, leading to the notion of ‘functional cat- egories’ and to a general conception of morphology as the ‘syntax of words’. These inflectional sub-word ‘units’ have long presented some of the most stubbornly recalcitrant challenges for morphemic analysis. Initial attempts to align individual inflectional features with morphotactic units created analytical conundrums in languages as inflectionally impoverished as English. Harris (1942: 113) and Hockett (1947: 240) struggled with the task of segmenting English children into morphemic units, due to uncertainty about the synchronic status of the historical strong (-r)andweak(-en) plural markers. Cognate patterns in other West Germanic languages raise similar problems. For example, some nouns in Modern German distinguish singulars and plurals by an ending, as illustrated by Tag~Tage ‘day(s)’. Other nouns mark the contrast by a medial vowel alternation as in Garten~Gärten ‘garden(s)’. In other nouns, the contrast is marked both by an ending and vowel alternation, as in Fuchs~Füchse ‘fox(es)’.Inyetother nouns, such as Kabel ‘cable(s)’ there is no variation between the singular and plural. A biunique correspondence cannot be established in a uniform manner between the feature ‘plural’ and a ‘unit of form’ in these cases without assigning more abstract analyses to the surface forms. Various technical strategies have been explored for assigning morphemic analyses to these and other seemingly non-biunique patterns of inflectional marking. One type of proposal generalizes the notion of ‘form’. Among the initial generalizations were different varieties of ‘special morphs’,suchasthe‘process morphs’ that bundled pairs of alternating vowels into ‘units’. Modern descend- ants include morphophonemic ‘readjustment rules’ (Halle and Marantz 1993), which intervene between morphemic analyses and surface forms. An alternative strategy involves reclassifying patterns of surface variation, so that some markers can be discounted in determining morphemic biuniqueness (Noyer 1992). Inflection at the morphology-syntax interface 233
Contemporary interest in exploring these types of technical refinements tends to be concentrated in communities with an overarching commitment to a syntacti- cocentric conception of morphology. From a morphological perspective, the core problems of morphemic analysis derive from an excessively narrow view of mor- phological structure and, as such, do not seem amenable to purely technical solu- tions. Instead, as argued in Matthews (1972, 1991), biunique relations are best understood as limiting cases of generally many-to-many relations between inflec- tional features and units of form. The extended discussion of Latin conjugational patterns in Matthews (1972) traces the problems of segmentation and interpreta- tion created by coercing a morphemic analysis onto languages that do not con- form to an agglutinative ideal. Many-to-many relations between features and forms are so endemic to Latin and Ancient Greek that they are robustly exhibited by regular and even exemplary items. To illustrate this point, Matthews (1991: 174) considers the Ancient Greek form elelýkete (e-le-ly-k-e-te) ‘you had unfas- tened’, which, as he notes, does not show “any crucial irregularity” and “is in fact the first that generations of schoolchildren used to commit to memory”:
But categories and formatives are in nothing like a one-to-one relation. That the word is Perfective is in part identified by the reduplication le- but also by the suffix -k-. At the same time, -k- is one of the formatives that help to identify the word as Active; another is -te which, however, also marks it as ‘2nd Plural’. (Matthews 1991: 173)
The deeper problem, as Matthews emphasizes, is not just that ‘flectional’ lan- guages like Latin or Greek appear to exhibit morphemic indeterminacy, but that the indeterminacy is the artifact of a method. By foisting an agglutinative analysis onto flectional languages, a morphemic approach creates problems for which it can provide no principled solution. The attempt to address these prob- lems through technical refinements of a morphemic model seems futile; at least some languages falsify the assumptions of the model:
One motive for the post-Bloomfieldian model consisted, that is to say, in a genuinely fac- tual assertion about language: namely, that there is some sort of matching between mini- mal ‘sames’ of ‘form’ (morphs) and ‘meaning’ (morphemes). Qua factual assertion this has subsequently proved false: for certain languages, such as Latin, the correspondence which was envisaged apparently does not exist ...One is bound to suspect, in the light of such a conclusion, that the model is in some sense wrong. (Matthews 1972: 124)
Subsequent studies have provided further confirmation that inflectional features and units of form are, in the general case, related by many-to-many ‘exponence’ relations. One strand of this research has even explored relations that are more ‘exuberantly’ many-to-many than the patterns exhibited by classical languages (Harris 2009; Caballero and Harris 2010). A pair of related conclusions can be 234 Claudia Marzi et al. drawn from this work. The first is that, when applied to all but the most uniformly agglutinative structures, morphemic models create problems of analysis while also obscuring the organization and function of form variation. The second is that mor- phemes cannot provide the basis for inflectional description and that even descrip- tive conventions like ‘morpheme glosses’ harbor untenable idealizations.
2.2 Varieties of exponence
Although exponence relations do not share the descriptive shortcomings of mor- phemes, they exhibit their own characteristic limitations. What unites the diverse exponence relations investigated in the realizational literature is a fundamentally negative property: the relations all involve non-biunique feature-form associa- tions. The realizational tradition offers no positive characterization of these rela- tions, and contains almost no discussion of what functions, if any, might be associated with different patterns of exponence. A model that recognizes many-to-many feature-form relations sacrifices the attractively simple compositional semiotics of a morphemic model, in which complex forms and complex meanings are built up in parallel from atomic form-meaning pairs. As expressed by the ‘Separation Hypothesis’ (Beard 1995), realizational accounts treat feature bundles as ‘minimum meaningful units’ and assign no meaning or function to variation in the ‘spell-out’ of bundles. In effect, realizational models move from one extreme to another, replacing indi- vidually meaningful morphemes with collectively meaningless exponents. Both of these extremes reflect a set of limiting assumptions about the nature of inflectional functions and meanings. Three assumptions are of particular im- portance. The first is that meanings are exclusively ‘extramorphological’ and do not encode information about the shape or distribution of related forms, or other properties of the morphological system. The second is that discrete meanings are associated statically with forms, not determined dynamically within a network of contrasts. The third is that analyses and interpretations are taken to be assign- able to forms in isolation from the systems in which they function. By incorporating these assumptions, realizational approaches close off any inquiry into the meaning or function of different patterns of inflectional expo- nence. The adoption of other, equally conservative, assumptions imposes simi- larly severe constraints. Whereas the role of paradigmatic structure has been a matter of dispute for most of the modern period, the relevance of morphotactic structure has remained almost unquestioned. However, in even the best-studied languages, there have been no systematic attempts to provide cognitive motiva- tion for the morphotactic structures assigned in standard descriptions. It is Inflection at the morphology-syntax interface 235 sometimes believed that a shift from orthographic to phonemic representation places analyses on a firmer foundation. But it is well-known that inflectional con- trasts may be cued by sub-phonemic contrasts (Baayen et al. 2003; Kemps et al. 2005). Moreover, the crude procedures of distributional analysis used to deter- mine morphotactic structure have no provision for distinguishing sequences that function as units in the synchronic system from those that are diachronic relics of the processes of morphologization that produced the system.
2.3 Synchronic vs diachronic structure
The confound between synchronically active structure and diachronic residue derives ultimately from the general conflation of synchronic and diachronic di- mensions in the post-Bloomfieldian tradition. To a large extent, this tradition is guided by the goal of exploring the descriptive potential of recasting diachronic analyses in synchronic terms. Underlying representations are transparent syn- chronic proxies for the ‘least common ancestors’ of a set of surface forms. Due to the regularity of sound changes, these ancestors will tend to be definable in terms of a minimum edit distance from their descendants. The serial structure of derivations likewise mirrors the temporal order of the sound changes that occur in the history of a language. The agglutinative bias of a morphemic model (termed ‘The Great Agglutinative Fraud’ by Hockett (1987: 83)) also re- flects an essentially historical perspective. A perfectly agglutinative system is one in which the waves of grammaticalization that build up complex forms pre- serve the discrete meaning and morphotactic separability of the morphologized parts, while maintaining a link between their meaning and form. However, the shift from a diachronic to a synchronic perspective is as dis- ruptive to the interpretation of a post-Bloomfieldian model as the move from biuniqueness to many-to-many exponence relations is for the semiotics of a realizational approach. Morphotactic structure is often a useful guide to the his- torical processes that applied to produce the inflected forms of a language. Yet knowledge of the historical origins of forms can distort analyses of their current status and function. This problem arises in an acute form in the ‘templatic’ analyses assigned to languages of the Algonquian and Athapaskan families. One tradition of analysis, illustrated in the Navajo grammar of Young and Morgan (1987), associates verbs with morphotactic ‘templates’ consisting of se- quences of ‘slots’, each containing a substitution class of formatives. An intri- cate overlay of distributional and interpretative dependencies holds between the ‘choices’ at different points in the template. Not all slots need be ‘filled’ in with a surface form and, typically, many are empty. As a result, there is a vast 236 Claudia Marzi et al. contrast between the complexity of the ‘underlying’ templatic structure and the morphotactics of surface forms. In effect, the dimensions of variation in the templatic descriptions provide a record of the history of the derivation of the modern forms. The apparent ‘choices’ are no longer independent in the modern languages but have been encapsulated in larger recurrent sequences, as repre- sented in the type of bipartite structure proposed by McDonough (2000). In sum, despite the appeal of a uniform approach to syntactic and inflec- tional patterns, the models developed within the broad post-Bloomfieldian tra- dition incorporate biases and assumptions that have mostly impeded attempts to understand the structure and organization of inflectional systems. The ana- lytical assumptions incorporated in approaches that were developed expressly to model inflectional patterns offer a usefully different perspective.
3 Paradigmatic distribution and interpretation
As their name suggests, classical ‘Word and Paradigm’ (WP) models shift the fun- damental part-whole relation in an inflectional system onto the relation between individual words and inflectional paradigms. This shift does not deny the signifi- cance of sub-word variation but instead interprets variation in a larger paradig- matic context. The context is defined by the interaction of two main relational dimensions. The first dimension is discriminative, implicit in the etymological origin of the term ‘inflection’, which, as Matthews (1991: 190) notes, derives “from a Latin Verb whose basic meaning was ‘to bend’ or ‘to modify’”.Patterns of form variation serve to distinguish larger ‘free forms’ with an independent dis- tribution that can communicate meaning or function within a system. In contrast to morphemic models, variants are not individually meaningful; in contrast to realizational models, they are not collectively meaningless. The variants that dis- criminate a form determine its place in a similarity space defined by the inflec- tional contrasts exhibited by a language. The second dimension is implicational or predictive, expressed by “the ... general insight ... that one inflection tends to predict another” (Matthews 1991: 197). Patterns of variation within an inflectional system tend to be interdepen- dent in ways that allow speakers to predict novel forms on the basis of forms that they have encountered. The predictive patterns that hold between forms deter- mine the place of a form within implicational networks. Discriminative and implicational relations are particularly relevant to the or- ganization of inflectional systems, which exhibit highly uniform patterns of con- trast and predictability. The inflected variants of an open-class item typically Inflection at the morphology-syntax interface 237 define a closed, uniform feature space and a largely transparent semantic space. For an item of a given word class or inflection class, it is possible to specify the features that are distinctive for that item. This degree of uniformity and item-in- dependence is what permits the strict separation of features and form in a reali- zational model. From even partial exposure to an inflectional system, speakers can arrive at reliable expectations about the number of forms of an item, assign interpretations to these forms, and even predict the shape of not yet encountered and potential variants. The discriminative and implicational organization of inflectional systems is saliently reflected in the ways that these systems are described. The notion of a ‘morphological gap’ tends to be applied to cases in which predictable inflected forms are either missing or occur less frequently than expected (compared with the frequency of corresponding forms of other items of the same class). Unattested derivational formations are more rarely described as ‘gaps’, since a derivational system is not conceptualized as a closed, fully populated, space of forms. The no- tions ‘suppletion’, and even ‘syncretism’, also apply almost exclusively to in- flected forms. Suppletion reflects unmet expectations about predictability and syncretism violates assumptions of discriminability. In the derivational domain, both patterns are usually treated as cases of ambiguity. Conversely, whereas deri- vational formations are often described as ‘established’, this term is rarely applied to individual inflected forms.
3.1 Words
The role of words and paradigms is a distinctive characteristic of classical WP descriptions of inflectional patterns. The use of word forms to exhibit variation accords with the view that “[t]he word is a more stable and solid focus of gram- matical relations than the component morpheme by itself” (Robins 1959: 128). In support of this claim, paradigmatic approaches present individual case studies but no systematic attempt to explain why words provide a useful basis for inflec- tional description. Instead, word-based analyses of inflectional systems exploit the generally positive correlation between unit size and grammatical determi- nacy. In part, this correlation derives from the monotonic nature of determinacy. A fully determinate form may be composed of individually indeterminate parts. To take a simple example, the indeterminacy associated with an element such as English -er, which may mark comparative forms of adjectives or agentive nomi- nals, is resolved in the word form spammer. In contrast, a set of fully determinate parts cannot be arranged into an indeterminate whole without contradicting the original assumption that they are determinate. 238 Claudia Marzi et al.
The modern development of a quantitative and cognitively-grounded dis- criminative approach offers a further perspective on this issue. As ‘free forms’ in the sense of Bloomfield (1933), words are the smallest units with an indepen- dently statable range of distributions and uses. As such, words are the smallest units with the kinds of individual ‘conditioning histories’ that are approximated by semantic vectors in models of distributional semantics such as Marelli and Baroni (2015). Correlating with these distributional and functional properties are measures of the uncertainty at word boundaries and the ‘informativeness’ of these boundaries, as operationalized by effects on compressability (Gertzen et al. 2016). The interpretive determinacy of words also underlies the ‘morphological in- formation’ that they express, i.e. information about the shape and distribution of other forms in a system. A form in isolation is of limited diagnostic value, reflect- ing the fact that such ‘pure’ forms are ecologically invalid abstractions. Speakers encounter forms in syntagmatic and paradigmatic contexts that guide their inter- pretation and support reliable inferences about related forms. The Paradigm Structure Constraints of Wurzel (1970) represent an early attempt to model this structure in terms of logical implication. Subsequent approaches, exploiting in- sights from the morphological processing tradition represented by Kostić et al. (2003) and Moscoso del Prado Martín et al. (2004), develop more robust informa- tion-theoretic measures of implicational structure. This approach has offered a useful perspective on questions concerning the structure and ‘complexity’ of pat- terns and classes, the nature of defectiveness and other properties of inflectional systems. Overall, the approach lends support to the traditional view that the in- flectional component is not an unstructured inventory but, rather, a system, in which patterns of interdependency facilitate predictions about the whole system from a subset of its forms. Patterns of predictability also contribute to an under- standing of the learnability of complex inflectional systems, and help to clarify thedegreeofvariationinthecomplexityofdescriptionsofinflectionalsys- tems. In at least some cases, this variation reflects the intrinsic difficulty of describing systems as assemblages of independent items that are composed of recurrent parts. Although this may provide a reasonable basis for enumer- ating the patterns attested in a language by means of a written grammar, it is not a plausible model of acquisition or use. Much of the extreme cross-linguistic variation in grammars and typological accounts appears to be an artefact of de- scriptive tasks that never arise for speakers and hence are not subject to pres- sures that ensure learnability. Inflection at the morphology-syntax interface 239
3.2 Paradigms
In descriptions of inflection class morphology, the variation exhibited by a class system is typically illustrated by full paradigms of ‘exemplary’ items, to- gether with diagnostic ‘principal parts’ for non-exemplary items. There has been a tendency for this descriptive practice to be overinterpreted, by advocates as well as by opponents of paradigmatic approaches. Reflecting the pedagogical origins of the classical tradition, exemplary para- digm and principal part descriptions are designed to provide a description of in- flectional patterns that is maximally transparent for learners. There is neither cognitive nor empirical motivation for assuming that this pedagogical organiza- tion mirrors the format in which speakers represent knowledge about inflectional patterns. This observation can be traced at least to Hockett (1967), in speculations about the divergence between pedagogical and cognitively-plausible models of paradigmatic structure, and morphological description in general:
in his analogizing ...[t]he native user of the language ...operates in terms of all sorts of internally stored paradigms, many of them doubtless only partial; and he may first en- counter a new basic verb in any of its inflected forms. (Hockett 1967: 221)
The intervening half century has seen the development of methodologies for probing speakers’ morphological knowledge, though the most robust measures, such as morphological family size (de Jong et al. 2000; Mulder et al. 2014), apply to derivational families. In parallel, large-scale corpora and other lexical resources have provided detailed information about the composition and struc- ture of the forms ‘in circulation’ within a speech community. It is well estab- lished that forms exhibit a Zipfian distribution (Zipf 1935), in which the frequency of a word is (approximately) inversely proportional to its rank in a frequency table. There are two particularly significant consequences for the learning of inflectional systems. First, roughly half of the forms that a speaker encounters will be occurrences of a small number of high frequency items (the, of, etc., in English). Second, almost half of the items of a language will be hapax legomena, i.e., forms that occur exactly once in a corpus, irrespective of its size, and thus are unlikely to be encountered multiple times by speakers. These distributional biases support Hockett’s conjecture about the nature of the input that a speaker encounters. They also argue against a naive pedagogi- cally-influenced conception of paradigmatic structure and against any version of the ‘Full Listing Hypothesis’ (Hankamer 1989) on which speakers would be as- sumed to memorize the inflected forms of a language. Speakers appear to encoun- ter only a small proportion of the forms of open-class items; for up to half of the lexicon, they may encounter just a single form. Nevertheless, speakers appear to 240 Claudia Marzi et al. be able to ‘pool’ the variation exhibited by partial paradigms and extend attested patterns to new forms. The format of this knowledge and the mechanisms that speakers use to amalgamate and extrapolate patterns are not yet well understood. However, there is evidence to suggest that speakers exploit lexical neighborhoods to define an analogical base that can bootstrap the process (Blevins et al. 2017). The morphological processing literature has also investigated a range of ef- fects in which paradigmatic structure appears to play a significant role. Among the most striking of these is the relative entropy effects first reported in Milin et al. (2009a, 2009b). These studies found that speakers in a lexical decision task were sensitive to the divergence between the frequency distribution of an item’s inflectional paradigm and that of its inflection class. To take a specific example, the more closely the distribution of the forms of the Serbian noun pla- nina ‘mountain’ matches the distribution of the ‘feminine’ declension in Serbian, the faster and more accurately speakers recognize forms of planina. This finding has been confirmed in follow-up studies including Baayen et al. (2011) and Milin et al. (2017).
3.3 Syntagmatic/paradigmatic integration
The interpretation of syntagmatic variation as, in part, a marker of paradigmatic relations also clarifies the motivation for the representational agnosticism of Neogrammarians such as Paul (1920). On the type of ‘constructive’ (Blevins 2006) interpretation of structure adopted in Post-Bloomfieldian models, forms are com- posedofadeterminatesequenceofatomicelements.Onamoreagnostic‘abstrac- tive’ interpretation, structure emerges from patterns exhibited by sets of forms. Multiple form classes are potentially relevant for the purposes of discriminating forms and deducing implications. Hence, the determination of syntagmatic struc- ture, like the enumeration of inflection classes, is task-dependent. The structure exhibited by forms, like the number of classes contained by a language, depends on the purposes for which one is assigning structure or counting classes. Current approaches to the modelling of inflectional patterns address a broad range of other issues. Network architectures, particularly ‘wide learning’ net- works (Baayen and Hendrix 2017) provide a representation of paradigmatic struc- ture that avoids the ‘combinatorial explosion’ (Baayen et al. 2013) associated with lists of forms, particularly if surface variants are encoded by exemplars. Models of vector semantics provide a way of grounding the interpretation of forms in an observable dimension of variation, i.e. distribution. The other observ- able dimension, variation in shape, is described by discriminable contrasts. Taken together, these dimensions offer an interpretation of the features Inflection at the morphology-syntax interface 241 employed in the description of inflectional systems, without assigning a ‘mean- ing’ to the features themselves.
4 Periphrasis and grammaticalization
We have to do with periphrasis if for certain cells of an inflectional paradigm no synthetic morphological form is available. Instead, a combination of words, an analytic or periphrastic form has to be used. For instance, Latin has no synthetic forms for the perfective passive of verbs, as illustrated here for the 3SG forms of the verb laudāre ‘to praise’:
Table 1: Imperfective and perfective 3SG forms of laudāre.
Imperfective Active Passive
PRESENT laudat laudātur PAST laudābat laudābātur FUTURE laudābit laudābitur
Perfective Active Passive
PRESENT laudāvit laudātus/a/um est PAST laudāverat laudātus/a/um erat FUTURE laudāverit laudātus/a/um erit
The cells for the perfective passive are a combination of the passive participle (that, like adjectives, agrees with the subject of the clause with respect to case, gender and number) and a form of the verb esse ‘to be’. If these word combina- tions were not considered part of the verbal paradigm, Latin verbs would have a paradigm with a gap for the perfective passive forms. These periphrastic forms receive a perfect interpretation, although the forms of the verb esse ‘to be’ are that of the imperfect tense. An additional argument for considering these word combinations as filling paradigm cells is the following. Latin has a number of so-called deponent verbs, verbs with a passive form but an active meaning. For instance, the verb lo- quor ‘speak’ is such a deponent verb. The crucial observation is that a word- sequence such as locutus est receives an active interpretation as well, and means ‘he has spoken’. This parallelism in interpretation as active meanings is to be expected if these analytic forms belong to the inflectional paradigm ofverbs(Börjarsetal.1997). 242 Claudia Marzi et al.
This means that phrasal constructions may express inflectional properties, just like morphological constructions (Ackerman and Stump 2004; Börjars et al. 1997; Sadler and Spencer 2001). Moreover, they cannot be analyzed as regular syntactic combinations because the morphosyntactic features of these constructions cannot always be determined on the basis of those of their word components (Ackerman and Stump 2004; Popova 2010). This can be illus- trated by the periphrastic constructions for perfect tense in Dutch. These are combinations of the verbs hebben ‘have’ or zijn ‘be’ with the past participles of verbs, as in:
(2) Jan heef-t het boek ge-lez-en John have-3S DET book PTCP-read-PTCP ‘John has read the book’
(3) Het meisje is ge-vall-en DET girl be.3S PTCP-fall-PTCP ‘The girl has fallen’
These two verbs are called auxiliaries, as they do not have their regular meaning in periphrasis, but instead express perfect aspect. The auxiliaries in these senten- ces have a present tense form. The past participles do not carry the perfective meaning either, as they can also be used with an imperfect meaning in passive sentences. Hence, the perfect meaning is the holistic property of the word combi- nation as a whole. This is why we speak of periphrastic constructions, because constructions are form-meaning combinations with possibly holistic properties that cannot be deduced from properties of their constituents. Periphrastic constructions require a formal analysis which does justice to these holistic properties. One model is that of Paradigm Function Morphology. In this model phrasal combinations can function as the realization of the set of morphosyntactic and/or morphosemantic features of a lexeme. For instance, heeft gelezen in example (2) is treated as the realization of the lexeme lezen ‘to read’ with the feature values [3.sg.perf] (Popova 2010). In the framework of Construction Morphology, periphrastic constructions are accounted for by means of the notion of ‘constructional idiom’ (Booij 2010). In English, for instance, the perfect tense form of verbs is a complex verbal predi- cate that consists of a form of the verb have with a past participle. This specific pattern expresses the perfect meaning, which cannot be derived from the mean- ing of one of the constituent words: neither the verb have, nor the participle itself is the carrier of the perfect meaning. The following constructional idiom may be assumed for the English perfect tense construction: Inflection at the morphology-syntax interface 243
(4) [[have]Vi [past participle]Vj]Vk ↔ [PERF [SEMj]]SEMk where PERF stands for the meaning ‘perfect tense’, and SEM for the meaning of the relevant indexed constituent. The use of auxiliaries for periphrastic constructions is a case of grammatic- alization, i.e. the process by which words with an original lexical meaning ac- quire a more grammatical meaning (Hopper and Traugott 1993). For instance, English have has the lexical meaning of ‘possess’ but this is not the relevant meaning in the periphrastic tense forms. Dutch has two auxiliaries for perfect tense, reflecting two different historical sources of these constructions, a posses- sive construction and a predicative one. In English the predicative construction was the source of the use of the auxiliary be, as in older English He is fallen ‘He is in a situation of fallen-ness’, which is now replaced in present-day English by He has fallen. This illustrates that English have is even more grammaticalized than its Dutch counterpart.
5 Lexicon and inflection in computational morphology
In computational terms, many aspects of the theoretical debate dealt with in the previous sections boil down to the problem of dealing with cases of many- to-many surface relations between form and function at the word level. Consider the representations in (5) and (6), where the Italian verb forms tengo ‘(I) hold’ (present indicative, 1st person singular) and tieni ‘(you) hold’ (present indicative, 2nd person singular) – from the verb TENERE ‘hold’–are segmented into their surface constituents.
(5) [[teng] HOLD +[o] pres ind 1s] HOLD pres ind 1s
(6) [[tien] HOLD +[i] pres ind 2s] HOLD pres ind 2s
The same word forms are split into more abstract sub-lexical constituents in (7) and (8):
(7) [[ten] HOLD +[o] pres ind 1s] HOLD pres ind 1s
(8) [[ten] HOLD +[i] pres ind 2s] HOLD pres ind 2s 244 Claudia Marzi et al.
Representations in (7) and (8) define a sign-based (biunique) relationship be- tween the stem ten and its lexeme HOLD. This is blurred in (5) and (6), where lexical formatives are not identical, leaving us with the problem of classifying teng and tien as stem allomorphs of the same verb. In the generative linguistic literature, shorter lexicons are generally preferred over more verbose ones, under the standard generative assumption that what can be computed should not be stored. However, for a fully-inflected form to be produced from the representa- tion in (7), some adjustment rules have to be added to the grammar (e.g. a velar insertion rule before back vowels in the case at hand). In the Italian conjugation, many of these rules do not apply across the board, but obtain for particular lex- emes in specific contexts only (Pirrelli and Battista 2000). So the question of de- scriptive economy cannot really be confined to the lexicon, since a compact lexicon may call for a rather profligate set of ad hoc rules. Information theory provides a means to address these empirical issues on a principled basis. In an information-theoretic adaptation of Harris’ ideas (1951), Goldsmith (2001, 2006) models the task of morphological induction as a data compression (or Minimum Description Length, MDL) problem: “find the battery of inflectional markers forming the shortest grammar that best fits the empirical evidence” (Rissanen 1989). The grammar is a set of paradigms, and the empirical evidence a reference corpus, where each form occurs with a specific frequency distribution. In this framework, MDL penalizes two descriptive extremes. First, it disfavors an ex- tremely redundant grammar, where each word form is part of a singleton para- digm that has that form as its only member. This in fact amounts to a repository of fully listed words, where each form is assigned the probability with which it occurs in the corpus. At the opposite end, a very compact grammar contains one overall paradigm only, where any verb stem can freely combine with any affix. This is a very short and redundancy-free grammar, but it overgenerates wildly, thus providing a poor approximation of the distributional evidence of the forms attested in the corpus. In what follows, we consider a rather more classical computational approach to this issue, using Finite State Transducers.
5.1 Finite State Transducers
A Finite State Transducer (FST) is an abstract computational device that turns an input string into an output string. Figure 1 depicts an FST for the present indica- tive stems of the Italian verb TENERE ‘hold’:namelyten-, tien-andteng-.Inthe graph, nodes (circles) are memory states, and directed arcs (arrows) represent transitions from one state to another. Each arc is decorated with either a single Inflection at the morphology-syntax interface 245 symbol, or a pair of symbols separated by a colon. The colon is a mapping opera- tor and reads: “the symbol on the left (or lexical symbol) is mapped onto the symbol on the right (either a surface symbol or a gloss)”. If only one symbol ap- pears on the arc, this means that the symbol is mapped onto itself, i.e. lexical and surface symbols are identical. Lower-case characters in the Latin alphabet represent simple letters; ‘ϵ’ is a meta-symbol representing the null character; and ‘Σ’ is a variable ranging over any input symbol. Finally, upper-case Latin charac- ters are linguistic glosses that express morphosyntactic features (e.g. ‘PAST’ or
‘3S’), stem indexes (e.g. ‘B1’)andlexicalcontent(‘TENERE’).
Figure 1: A Finite State Transducer for the present indicative stem allomorphs of Italian TENERE ‘hold’.
Starting from the initial state q0, the FST reads one input symbol at a time from left to right, until it consumes the whole input string. Upon reading a symbol (e.g. ‘t’ in teng-) in a particular state, the transducer looks for a state- leaving arc that is labelled with the same input symbol. If such an arc exists, it is traversed, and the mapping operation annotated on the arc is carried out. An input string is successfully processed if the FST finds itself in the final state f after reading the final letter of the input string. Note that all branches of the FST in Figure 1 end up with a stem index (Bi)beinginserted:e.g. ‘ϵ:+B2’. The index enforces the constraint that the just read off stem allo- morph is followed by an inflectional ending selecting the appropriate index.1
1 Stem indexing (Stump 2001) enforces a relationship between a verb stem and a set of cells in the verb paradigm. Although stem indexes and stem formation processes are often mutually implied (as illustrated by the FST in Figure 1), they might occasionally be independent (Pirrelli & Battista 2000). 246 Claudia Marzi et al.
For example, in the present indicative, teng- canbefollowedonlybythefirst singular and third plural ending. A number of properties make FSTs theoretically interesting devices. FSTs are not classical rewrite rules. They enforce a correspondence relation between two distinct levels of representation. The relation is declarative, parallel and bi- directional. We can easily swap input and output symbols and use, for word generation, the same transducer designed for word parsing. This means that we can understand the teng-ten relation in either of the following ways: (i) g is trailed after ten- to incrementally (in Stump’s 2001 terms) add B2 to the result- ing representation; (ii) teng is the result of applying to ten a specific paradigm function that realizes the appropriately indexed allomorph B2. In the former case B2 adds some (disjunctive) morphological content. In the latter case, it is only the indexed trace of a realizational process. Secondly, FSTs represent lexical items as a union set of intersecting autom- ata. Since automata define processing operations over lexical and surface repre- sentations, the whole finite state machinery amounts to a two-level word graph incorporating lexical information, morphotactic constraints and context-sensitive formal changes, all cast into a uniform format. FSTs thus blur the distinction between lexical representations and rule schemata defined over lexical represen- tations. This makes them extremely powerful processing devices, capturing mor- phological generalizations at fine-grained levels of generality, ranging from sweeping, exceptionless phonological processes to lexically-conditioned surface adjustments. The overall emerging view is closer to the idea of a large, integrated repository of both specific and general information, than to the classical notion of a generative grammar consisting of a lexical repository and a set of rules. On a less positive note, the high expressive power of FSTs can get a major liability. For a lexicon of average inflectional complexity, it soon becomes ex- tremely cumbersome to craft by hand the whole range of fine-grained mapping relations that are needed for recognition/production. And this can be uselessly laborious. For example, since every stem must be associated with a specific cell index, transitions to cell-specific indexes must also be stipulated for the single- ton stem of a regular paradigm. We thus miss the obvious generalization that only irregular paradigms require explicit stipulation of the transitions from stem alternants to specific subsets of paradigm cells. The DATR formalism (Evans and Gazdar 1996) provides an effective framework for implementing de- fault statements through inheritance relations in the lexicon, and can be used to enforce both high- and low-order constraints on stem allomorphy. Inflection at the morphology-syntax interface 247
5.2 Hierarchical lexicons
A DATR lexicon is an inheritance network of nodes, each consisting of a set of morphologically related forms, ranging from major inflection classes (Table 2, columns 1 and 2), to regular (Table 2, column 3) and subregular paradigms (Table 2, column 4). An inflected form like, e.g., tengo is obtained from the ab- stract lexical schema ‘X+o’ by assigning ‘X’ an appropriate stem. Each entry node is defined by a “name” (the unique identifier in bold ended by a colon), followed by a few “statements”, each taking one line in Table 2, with the following general format: “attribute == value” (where ‘==’ is an assignment operator). An attribute is surrounded by angle brackets. It can consist of a single label (e.g. “”, in Table 2), or by a sequence of blank-separated labels, defining an increasingly specific attribute (e.g. “”). The empty label ‘<>’ denotes the most general or unspecified attribute. Any label added to ‘<>’ defines a branching node in the hierarchy: the further down we go in the hierarchy, the more specific the attribute. Accordingly, ‘
Table 2: DATR lexical entries for Second Conjugation (2C) Italian verbs.
Morpholex-paradigm-: C_VERB: TEMERE: TENERE: == “
Hierarchies of specificity play a crucial role in DATR information flow. General information, defined by attributes higher up in the hierarchy, tends by default to percolate to lower (more specific) attributes, unless the latter ex- plicitly contain overriding information. For example, the node ‘Morpholex- paradigm-0’ (column 1, Table 2) states that the index ‘B’ is assigned whatever string is assigned to the ‘root’. This information is inherited by the node ‘2C_VERB’ (short for “second conjugation verb”: column 2, Table 2) through the statement “<>== Morpholex_paradigm_0”,whichreads:“whatever 248 Claudia Marzi et al. information is in Morpholex_paradigm_0 is copied here”. This statement boils down to adding all statements in ‘Morpholex_paradigm_0’ to ‘2C_VERB’.At this level, however, ‘root’ is assigned no surface string. This is done in the node ‘TEMERE’ (column 3, Table 2), which stipulates that all statements of ‘2C_VERB’ (“<>== 2C_VERB”) are inherited locally: here, the attribute ‘root’ is instantiated with the string tem.Inturn,‘B’ is assigned the local value of ‘root’ (= “
Upon reflection, this is not just an issue of descriptive accuracy. In fact, it boils down to the deeper, explanatory question of how children can come up with the relevant knowledge needed to optimally process inflected forms. According to this learning-based view, redundant patterns are predominantly statistical, and even irregularities appear to be motivated by their frequency distribution in the system and the general-purpose learning strategies of the human language pro- cessor. All of this can admittedly be very different in character from the formal constraints on units, representations or rule systems proposed within theoretical and computational models. Nonetheless, it represents, in our view, an extremely insightful entry point to the grand issue of language architecture. In the following section we shortly consider some implications of the regular vs. irregular distinc- tion from a developmental perspective, to then move on to the computational modelling of inflection acquisition (see Ravid, Keuleers and Dressler 2020, this volume, for a more comprehensive overview of morphology acquisition).
6 Acquiring inflection
6.1 The logical problem of acquiring inflection
According to a classic account (Berwick 1985; Pinker 1989, 1984), the task of a child attempting to learn how verbs and nouns are inflected can be described as involving grammar hypothesis testing. This is not too dissimilar from the MDL grammar evaluation framework illustrated in the previous section. Children need to arrive at a grammar hypothesis H that includes all well-formed inflected forms of the input language, while excluding ill-formed ones. An important logical problem with this account is represented by the case when the child’sgrammar is a superset of the target grammar (Pinker 1989): the child masters all correctly inflected forms, but nonetheless also produces some ungrammatical forms (e.g. (s)he says went and *goed interchangeably). How can children recover from these errors? In the acquisitional literature, the question of whether children can correct themselves on the basis of received explicit negative evidence (e.g. negative correction by their care-givers) has been highly debated. Some scholars suggest there is little reason to believe that children are supervised in this way (Bruck and Ceci 1999; Taatgen and Anderson 2002; but see Chouinard and Clark 2003 for a different view). Even when they are corrected (which is neither frequent nor systematic), they may take little notice of correction. In contrast with this position, other scholars (Kilani-Schoch et al. 2009; Xanthos et al. 2011) 250 Claudia Marzi et al. emphasize the impact on child’s morphology learning of both explicit and im- plicit, positive and negative feedback by parents, thereby questioning specula- tive claims (of direct or indirect Chomskyan inspiration) of poor and noisy input evidence in child-directed speech. Proponents of lack of corrective feedback in child’s input have hypothe- sized that possibly innate mechanisms for self-correction are in place. For ex- ample, according to Marcus et al. (1992), a blocking mechanism may suppress ed-verb past formation in English *goed, due to the stored representation of ir- regular went being entrenched in the lexicon by repeated exposure. This is in line with Pinker and Ullman’s (2002) ‘Words and Rules’ theory, according to which only irregularly inflected forms are stored in full. Regulars are either combined online from their stems and affixes in word generation, or are split into stems and affixes in recognition under the assumption that their access units are sublexical. In producing an inflected form, the lexicon is accessed first, and on-line assembly is pre-empted if the target (irregular) form is found there. Otherwise, a regular form is produced by using combinatorial rules. However logically elegant and simple, lexical blocking seems to make rather unrealistic assumptions about how children come up with a regular vs. irregular distinction while being exposed to inflectional systems of average complexity. For example, Pinker and Ullman (2002) suggest that there is only one default regular pattern in the inflection of any language, and that the deci- sion is dichotomous: the child considerably constrains the hypothesis space, as only one inflectional process can be held as a “regular” candidate for any spe- cific paradigm function. This assumption, however, appears to seriously under- estimate the systemic complexity of highly-inflecting languages. For example, the verb conjugations of Italian and Modern Greek show a graded hierarchy of regularity-by-transparency effects of morphological processing: many verbs present phonologically predictable adjustments that obscure the stem-ending boundary; some others keep the transparency of the stem-ending boundary at the price of introducing formally unpredictable fillers like thematic vowels, etc. Thus, the central question is how children can possibly home in on the decision of storing an irregular form as an unsegmented access unit, and a regular form as consisting of at least two distinct access units. We know that, for some mod- els of lexical access, the decision does not have to be yes-or-no. For example, so-called “race models” (Schreuder and Baayen 1995; Baayen et al. 1997) as- sume that words can possibly be stored as both whole forms and sublexical ac- cess units, thus providing two parallel access routes that are concurrently activated and adjudicated on the basis of frequency and task-based effects. However, the same models leave the issue of word segmentation seriously underspecified: how does the child segment a morphologically complex word Inflection at the morphology-syntax interface 251 that is taken to be regular? How and when does perception of sublexical struc- ture develop in child lexical competence to trigger lexical self-organization? In a comprehensive comparison of the developmental stages in the acquisi- tion of verb inflection in nearly two dozen languages (the Indo-European, Ugro- Finnic and Semitic families plus Turkish), Bittner, Dressler and Kilani-Schoch (2003) observe that the transition from rote lexical storage to morphological processing is the result of a process of active knowledge construction by the child, crucially conditioned by typological factors such as richness, uniformity and transparency of inflectional paradigms (Dressler 2010). Nevertheless, schol- ars widely differ in the way they conceptualize this process. In the framework of Natural Morphology (Dressler 2010, 2005), an increase in children’s inflectional productivity follows the establishment of the first “mini- paradigms”: i.e. non-isolated mini-sets of minimally three phonologically unam- biguous and distinct inflected forms of the same lemma. Miniparadigms thus mark the turning point between a “premorphological” and a “protomorphologi- cal” phase in child acquisition of inflection. The onset of this transition phase can vary considerably, depending on type and token frequency, lexicon size, phonological salience, regularity and transparency of the morphological system. Legate and Yang (2007) model child acquisition of English inflection as a maturation period during which the child starts entertaining a grammar hypothe- sis compatible with a language that does not manifest tense marking (e.g. Chinese), to eventually eliminate it. The crucial observation here is that elimina- tion of this hypothesis is a gradual process. The frequency of inflectionally un- marked usages by the child goes down incrementally, mostly over a 2–3-year span (Haegeman 1995; Phillips 1995). More interestingly for our present concerns, the length of such a maturational period and the frequency of inflectionally un- marked usages are influenced by the richness and complexity of the inflection system being acquired. Based on distributional evidence in the child-directed por- tion of the Brown’s (1973) Harvard Studies, the Leveille corpus and the Geneva corpus in the CHILDES database (MacWhinney 1995), Legate and Yang observe that the percentage difference between clauses with overt tense morphology (e.g. I walked to school) and clauses with no tense morphology (e.g. I make him run)is nearly 6% in English, 40% in French, and 60% in Spanish. Accordingly, they make the prediction that the maturational period for a child to acquire contextu- ally appropriate tense marking will be longer for English and shorter for Spanish. This is borne out by developmental data. The greater balance of inflectional con- trast in the Spanish verb system shows an inverse correlation with the lower per- centage of unmarked usages observed in Spanish children (17%), compared with German children (58%) and English children (87%) (data form the CHILDES cor- pus, reported in Freudenthal et al. 2010). 252 Claudia Marzi et al.
In a series of influential papers on discriminative language learning, Ramscar and colleagues (Ramscar and Yarlett 2007, Ramscar and Dye 2011) focus on the logical problem of acquisition of English noun inflection to show that chil- dren can in fact recover from superset inflection grammars by learning proba- bilistically-cued responses, according to the Rescorla-Wagner model of classical conditioning (Rescorla and Wagner 1972). Early in training, children go through a transitional phase where more forms like – say – *mouses and mice are pro- duced interchangeably. Due to the high number of -s plural nouns, compared with the relatively small frequency of mice, the pressure for mice to be replaced by the over-regularized *mouses is strong and dominates competition. In the end, however, children’s over-regularized response is surpassed by the “imita- tion” signal, i.e. by the increasing strength of the imitative response mice when its performance curve reaches asymptote. The discriminative account dispenses with both innate blocking mechanisms and parameterized grammar hypotheses. In the acquisition of verb inflection, input evidence of tense-marked forms like walked, ate and sang strengthens the associa- tion with the implicit morpho-semantic notion of [+Past] and the referential content of the three forms (namely WALK, EAT and SING). Root infinitives of the same verbs, on the other hand, will in turn reinforce associative links with [ − Past], and, over again, with WALK, EAT and SING. Thus, at early stages of English acquisition, the lexical semantics of each verb is more likely to be associated with verb forms that are unmarked for tense than with tensed forms (in line with Legate and Yang’s distributional evidence). This intra-paradigmatic competition (i.e. competition be- tween finite and nonfinite forms belonging to the same verb paradigm) turns out to be more balanced in a richer and more complex inflection system, like Spanish or Italian conjugation, where paradigm cells are, in the vast majority of cases, associ- ated with formally distinct, fully contrastive verb forms.
6.2 Paradigms and entropy
It is useful to see how word competition is related to the entropy HPðÞof the distribution of phonologically distinct members in the verb paradigm P,or “contrastive” paradigm entropy, defined in equation (1): X HPðÞ= − pfðÞi log2ðÞpfðÞi (1) fi2P where fi ranges over all formally distinct inflected forms of P, and the probabil- ity pfðÞi is estimatedP by the ratio between the token frequency tfðÞi and P cumu- lative frequency tfðÞi . fi2P Inflection at the morphology-syntax interface 253
Table 3 shows a (fictitious) distribution of inflected verb forms in three pres- ent indicative (sub)paradigms for English sing,Germansingen ‘sing’,and Spanish cantar ‘sing’, under the simplifying assumption that distributions do not vary across languages, but only depend on lexical information (the verb being inflected) and morphosyntactic information (the selected paradigm-cell). When we calculate the entropy of each present indicative paradigm based on its distinct inflected forms, we are assessing how uniformly distributed the formal contrast is within the paradigm. Even if we assume that frequency distributions do not vary across languages, the comparative complexity2 of the inflection sys- tem in each language appears to affect the amount of intra-paradigmatic formal competition: the more discriminable the inflected forms are, the more entropic (i.e. the more balanced) their competition will be. When more cells are assigned the same form, cumulative frequency by form winnows down contrastive para- digm entropy, prompting a greater bias for fewer forms. Accordingly, the English bare form sing is largely dominant in its own present indicative paradigm.3
Table 3: Frequency distributions of verb forms in the present indicative paradigms of English, Spanish and German; p(pc|s) is the conditional probability of a paradigm cell pc given a verb stem s. pres ind English Spanish German p-cell sing freq p(pc|s) cantar freq p(pc|s) singen freq p(pc|s)
S sing . canto . singe . S sing . cantas . singst . S sings . canta . singt . P sing . cantamos . singen . P sing . cantáis . singt . P sing . cantan . singen . Σ Σ Σ
2 The term “complexity” is used here in the intuitive sense of “formal variety/richness”.We will provide a more rigorous definition in Section 8 of this chapter. 3 Inthecaseofthesing present indicative subparadigm, contrastive entropy is calculated as follows:
− pðÞ sing · log2ðÞpðÞ sing − pðÞ sings · log2ðÞpðÞ sings = 0.87, where pðÞ sing and pðÞ sings are estimated as the ratio between their respective token frequencies in the subparadigm (25 and 10), and their cumulative token frequency (35). For the German and Spanish subparadigms, the same formula yields higher entropy values, respectively 1.68 and 2.28. 254 Claudia Marzi et al.
An important difference between Legate-Yang’s model and the naive discrimina- tive approach of Ramscar and colleagues is that, unlike the latter, the former makes the finite vs. non-finite competition hardly sensitive to lexical patterning in the data: any usage of finite marking in the child’s input will reward a [ + Tense] grammar across the board, lending support to any prospective tense-marking usage, irrespective of whether it involves an auxiliary, a copula or a lexical verb. Legate and Yang’s model assumes that children are not learning how to inflect words, but how to reject a particular grammar hypothesis (or parameter setting). Contrariwise, a discriminative model makes any further usage of tense marking contingent upon the amount of contrastive competition within aspecificparadigm. The prediction that paradigm-specific distributions of finite and nonfinite forms play a role in determining the time course of inflection development is supported by the pattern of results reported in four quantitative analyses of early child language (Hamann and Plunkett 1998; Krajewski et al. 2012; Pine et al. 2008; Wilson 2003). In particular, by counting the number of times three classes of inflected verb forms (1SG and 3SG copula BE, 1SG and 3SG auxiliary BE, and 3SG -s) are overtly realized by English speaking children in obligatory contexts relative to all such contexts (or production rate), a few key findings are reported. First, there are significant differences in the rate at which typically developing children produce copula BE (e.g. It’s good), auxiliary BE (e.g. I’m eating) and third person singular present forms (e.g. He runs), as indicated by the following ranking: cop BE > aux BE > 3SG. The second finding is that is is produced at a higher rate with pronominal subjects than with lexical subjects. Thirdly, both copula and auxiliary is are produced at a significantly higher rate than am; among all instances of is, those preceded by it are realized signifi- cantly more often than those preceded by he.
6.3 Form and distribution effects
Discriminative learning paves the way to a coherent operationalization of inflec- tion acquisition in terms of learning contrasts within and across the two observ- able dimensions of variation in an inflectional system (Blevins 2016): form classes and frequency distribution classes. The discovery of phonological and semantic sublexical invariants (e.g. stems in lexical paradigms, or inflectional endings in conjugation classes), has often been invoked as a mechanism accounting for the acquisition of inflectional morphology (Bybee 1988, 1995; Peters 1997; Penke 2012). Furthermore, an explanatory mechanism that crucially rests on the amount of competing sources of discriminative information in the input can naturally be ex- tended beyond the word level. To illustrate, the following nuclear sentences form a Inflection at the morphology-syntax interface 255 sort of multi-word or syntactic paradigm,4 defined as a set of constructions in com- plementary distribution and mutual competition.
(9) She walks to school She drinks cola She’s walking to school Does she drink cola?
From this perspective, “[g]rammatical features, properties and categories can likewise be interpreted as proxies for form classes, distribution classes or some combination of the two. In this way, morphological terminology that mislead- ingly implies an associated semantics can be reduced to robustly observable dimensions of form variation” (Blevins 2016: 249). Prima facie sign-based rela- tionships between form and content like 3SG -s in English verb conjugation are derived from the relation between a morphological marker and its embedding syntactic context. From this perspective, inflectional paradigms are descriptively useful shorthands for construction-based paradigms, where discriminative rela- tions are expressed between fully spelled-out lexical forms and grammatical forms, rather than between word forms and abstract paradigm cells. TheideaisinlinewithHarris’s (1968) distributional approach to word seg- mentation and morpheme segmentation, where structural boundaries are identi- fied at the points of likelihood discontinuity between adjacent symbols. We will consider these issues in more detail in Section 7.2.2. Note that a distributional, dis- criminative approach to inflection in fact conflates the two interlocked issues of how inflection is marked and in what contexts (see Section 1 above). Both aspects appear to capture distributional relations, and can be viewed as the same linguistic phenomenon looked at on different time scales. The distributional constraints that require verb marking, or agreement between two co-occurring units, are captured within a relatively large temporal window. Realizational effects are perceived on a shorter time scale. Contexts, as well as complex words, are split at their “joints” by contrasting “minimal” pairsofneighbors.Accordingly,paradigmaticrelations emerge as contrastive points in a multidimensional space where shared lexical units (e.g. she WALKs vs. he WALKed) or shared grammatical units (e.g. SHE IS walkING vs. SHE IS singING) are observed to occur in complementary distribution.
4 We can trace back the first formulation of this idea to P.H. Matthews’ Syntax (1981: 265–291), where a prescient construction-based interpretation of Chomskyan transformational rules is offered. 256 Claudia Marzi et al.
This view is supported by evidence that children acquire inflected forms by binding them to larger unanalyzed word “chunks” (Cazden 1968; MacWhinney 1976). Further support comes from evidence of children with Specific Language Impairment5 (SLI), who have difficulty with the interpretation of a complex sen- tence like The cow sees the horse eating (Leonard and Deevy 2011; Leonard et al. 2013). In addition, Purdy et al. (2014) show that older children with a history of SLI present a P6006 for subject-verb agreement violations as in *Every night they talkS on the phone, but not for violations as in *He makes the quiet boy talkS a little louder. If the children fail to grasp the structural dependencies between He makes and the clause that followed, no error would be detected, because the se- quence the quiet boy talks a little louder would seem grammatical. This data makes an interesting connection with the most significant proc- essing limitations observed in children with SLI: slow processing speed (Stark and Montgomery 1995; Kail and Leonard 1986; Wulfeck and Bates 1995, among others) and limited phonological working memory capacity (Briscoe et al. 2001; Farmer 2000; Montgomery 2004, among others). In particular, if children with SLI are weak in retaining the phonological representation of a new word (for long enough for it to be stored in long term memory), this may result in a lim- ited lexicon and a limited grammar. Development in the processing and pro- duction of complex sentences may also be delayed as a consequence of an impoverished lexicon, since complex sentences often involve verbs subcatego- rizing for sentential arguments. A more direct consequence of a limited working memory capacity in children with SLI is the difficulty to retain distant depen- dency relations in context, which explains their apparent insensitivity to viola- tions as in *He makes the quiet boy talkS a little louder. To sum up, inflection appears to develop in children as a highly interactive system, interfacing formal and functional features on different time scales. A rich, contrastive morphological system is helpful to acquire syntactic dependen- cies. At the same time, full mastery of inflection is contingent upon the intake of larger and larger syntactic contexts, where functional dependencies are realized as extended, multi-word exponents. Such a two-way implication is hardly
5 Specific Language Impairment is a significant deficit in language ability that cannot be at- tributed to hearing loss, low non-verbal intelligence, or neurological damage (Leonard 2014; Montgomery & Leonard 1998; Rice et al. 2000). 6 P600 is a peak in electrical brain activity that is measured with electroencephalography around 600 milliseconds after the stimulus that elicits it. P600 is commonly associated with hearing or reading grammatical (in particular, syntactic) errors (see Marangolo & Papagno 2020, this volume). Inflection at the morphology-syntax interface 257 surprising upon reflection, since different forms of the same verb paradigm are less functional if no verb arguments are yet expressed in context (Gillis 2003). We emphasized the largely distributional character of inflectional para- digms, whose cells define grammatical abstractions over a multi-dimensional network of formal contrasts in syntactic and pragmatic contexts. Nonetheless, paradigms acquire an autonomous relevance in word processing: they stake out the linguistic space where lexical forms get co-activated and compete in word recognition and production through contrastive formal oppositions. Contrastive paradigm entropy correlates with inflection rates in child pro- duction, and marks an important typological difference between inflecting lan- guages like Spanish or Italian and languages of the nearly isolating type such as English. Finally, as we will see in more detail in the ensuing sections, issues of intra-paradigmatic formal contrast are also relevant for understanding the communicative function of the opposition between regularly and irregularly in- flected words.
7 Machine learning of inflection
The task of modelling, with a computer, the dynamic process whereby a child gets to acquire her/his full morphological competence is reminiscent of Zellig Harris’ empiricist goal of developing linguistic analyses on the basis of purely formal, algorithmic manipulations of raw input data: so called “discovery pro- cedures” (Harris 1951). Absence of classificatory information (e.g. morphosyn- tactic or lexical information) in the training data qualifies the discovery algorithm as unsupervised. Conversely, when input word forms are associated with output information of some kind, then discovery is said to be supervised, and the task is modelled as a classification problem.
7.1 Constructive vs. abstractive approaches
Borrowing Blevins’ (2006) terminology, a useful distinction can be made here between constructive and abstractive algorithms for word learning. Constructive algorithms assume that classificatory information is morpheme-based. Word forms are segmented into morphemes for training, and a classifier must learn to apply morpheme segmentation to novel forms after training. An abstractive learning algorithm, on the other hand, sees morphological structure as emerging from full forms, be they annotated with classificatory information (supervised 258 Claudia Marzi et al. mode) or not (unsupervised mode). From this perspective, training data consist of unsegmented word forms (strings of either letters or sounds), possibly coupled with their lexical and morphosyntactic content. Accordingly, morphological learning boils down to acquiring knowledge from lexical representations in train- ing, to generalize it to unknown forms. In this process, word-internal constitu- ents can possibly emerge, either as a result of the formal redundancy of raw input data (unsupervised mode), or as a by-product of form-content mappings (supervised mode). In fact, for too many languages, morpheme segmentation is not a well- defined task, due to the notorious problems with the Bloomfieldian, sign- based notion of morpheme, and the non-segmental processes of introflexive (i.e. root and pattern), tonal and apophony-based morphologies. So, the as- sumption that any word form can uniquely and consistently be segmented into morpheme-like constituents is at best dubious, and cannot be enter- tained as a general bootstrapping hypothesis for morphology learning. In some abstractive algorithms, discovery procedures for morphological structure are constrained by a-priori assumptions about the morphology of the language to be learned. For example, knowledge that the target language mor- phology is concatenative biases the algorithm hypothesis search for stem-ending patterns. Thus, although no explicit morpheme segmentation is provided in training, the way word forms are tentatively split into internal constituents pre- supposes considerable information about boundary relations between such con- stituents (e.g. Goldsmith 2001). In some other algorithms, an alignment between morphologically related forms is enforced by either (i) using fixed-length posi- tional templates (e.g. Keuleers and Daelemans 2007; Plunkett and Juola 1999), or (ii) tying individual symbols (letters or sounds) to specific positions in the input representation (so-called “conjunctive” coding: Coltheart et al. 2001; Harm and Seidenberg 1999; McClelland and Rumelhart 1981; Perry et al. 2007; Plaut et al. 1996), or (iii) resorting to some language-specific alignment algorithms (Albright 2002) or head-and-tail splitting procedures (Pirrelli and Yvon 1999). However, the ability to recognize position-independent patterns in symbolic time series, like the word book in handbook, or the Arabic verb root shared by kataba ‘he wrote’ and yaktubu ‘he writes’, appears to lie at the heart of human learning of inflec- tion. Hence, a principled algorithm for morphological bootstrapping should be endowed with a capacity to adapt itself to the morphological structure of the tar- get language, rather than with a language-specific bias. In “features and classes” approaches (De Pauw and Wagacha 2007; McNamee and Mayfield 2007), a word form is represented as a set of redundantly specified n-grams, i.e. possibly overlapping substrings of n characters making up the input string: for example, ‘wa’, ‘al’,and‘lk’ for the string walk. N-grams have no Inflection at the morphology-syntax interface 259 internal structure and may be order-independent. The algorithm may start with the hypothesis that each word form is in a class of its own, and uses a stochastic classifier to calculate the conditional probability of having a certain class (a word form) given the set of distributed n-grams associated with the class. N-grams that occur in many words will be poorly discriminative, whereas features that happen to be repeatedly associated with a few word forms only will be given a morpho- logically meaningful interpretation. Discriminative approaches to learning such as “features and classes” have a lot to offer. First, they are able to deal with the problem of learning “a-mor- phous” morphologies on a principled basis, addressing traditional conundrums in the morpheme-based literature such as morphemes with no meanings (“empty morphemes”), meanings with no morphemes (“zero morphemes”), bracketing paradoxes, etc. Secondly, they seem to exploit the time-honored lin- guistic principle of contrast (Clark 1987, 1990), according to which any formal opposition can be used to mark a grammatical or lexical opposition. Thirdly, they bring word learning down to more general mathematical models of classi- cal conditioning (Rescorla and Wagner 1972) in behavioral psychology, accord- ing to which cues are constantly in competition for their predictive value for a given outcome. In what follows we will focus on abstractive, discriminative ap- proaches to word learning and explore their implications for models of inflec- tion. Such approaches see inflectional morphologies as complex adaptive systems, whose internal organization is the dynamic, continuously changing outcome of the interaction of distributional properties of input data, levels of lexical representation, and innate learning and processing constraints.
7.2 Associative vs. discriminative approaches
It is useful to describe discriminative learning by contrasting it with classical mod- els of associative learning. The gradient descent training of connections from input nodes to output nodes in a two-layer perceptron for word production is a well- known example of associative learning. In learning inflection, the task is to map an input representation (say “go PAST”) onto the corresponding output representa- tion (went). With PDP connectionist networks (Rumelhart and McClelland 1986), input representations are “wired in” on the input layer through dedicated nodes. A letter string like #go# (where ‘#’ marks the start and the end of the string) is simul- taneously input through context-sensitive, conjunctive encoding of each symbol together with its embedding context. Accordingly, the input g in #go# is encoded as a #_g_o node. Output representations are learned by adjusting connection weights through back-propagation of the error feedback. Back-propagation consists 260 Claudia Marzi et al. in altering the weights of connections emanating from the activated input node(s), for the level of activation of output nodes to be attuned to the expected output. According to the “delta rule” (equation (2)), connections between the jth input node and the ith output node are in fact changed in proportion to the difference ^ th between the target activation value hi of the i output node and the actually ob- served output value hi: ^ Δwi, j = γ · hi − hi · xj (2)
th th where wi, j is the weight on the connection from the j input node to the i out- th put node, γ is the network learning rate, and xj is the activation of the j input node. Note that, for xj = 0, the resulting Δwi, j is null. In other words, nothing changes in the connections emanating from an input node if that node is not activated. The “delta rule” in equation (2) is primarily associative. Learning proceeds by increasingly associating co-occurring cues (input nodes) and outcomes (out- put nodes), or by dissociating them when explicit evidence to the contrary is pro- vided by means of correction. Activation of an output node in the absence of an activated input node does not affect the connection from the latter to the former, and nothing is learned about their cue-response relationship. Interposition of a hidden layer of nodes mediating input and output nodes makes the relationship between input and output representations non-linear, but does not make it up for lack of explicit negative evidence. In the end, learning is based on the funda- mental assumption that the network is systematically “corrected”,i.e.itistold that an input-output connection is right or wrong. A more realistic connectionist approach to language learning is offered by Recurrent Neural Networks (hereafter RNNs), which provide a principled solu- tion to the problem of learning words with no external feedback (Elman 1990; Jordan 1997). RNNs are multi-layer perceptrons equipped with a first layer of hidden nodes interposed between the input layer and the output layer, and an extra layer of hidden nodes containing a copy of the activation state of the hid- den layer at the previous time tick. In RNNs, a string is input as a time series of symbols (not as a synchronous activation pattern), with each symbol being pre- sented at a discrete time tick. Furthermore, word learning is conceptualized as a prediction-driven task. Upon presentation of an individual symbol on the input layer, an RNN must guess the upcoming symbol on the output layer. In a nutshell, the network learns to predict sequences of symbols based on its past experience. This is fairly ecological: prediction is known to be heavily involved in human language processing (Altmann and Kamide 2007; DeLong, Urbach, and Kutas 2005; Pickering and Garrod 2007). In addition, it provides a natural Inflection at the morphology-syntax interface 261 answer to the lack-of-feedback problem. Everything the network has to do is to wait for an upcoming symbol to show up. If the symbol does not match the cur- rent network prediction, connection weights are adjusted for the current input symbol to be the most likely network’s response when a similar sequence is pre- sented over again. Simple RNNs are serial memories. They can improve on sequence predic- tion by keeping memory of the immediately preceding context through patterns of re-entrant connections in the hidden layer. At the same time, they plausibly address the problem of context-sensitive encoding of input symbols. The way a symbol is internally encoded by a RNN crucially depends on the network’s memory of embedding sequences where the symbol was found. Different train- ing data will yield different internal representations. There is no need for pre- wired input nodes encoding specific symbols in specific contexts. In principle, a RNN trained on strings from one language, can be trained on strings from an- other language, gradually adjusting its input nodes for them to be able to cap- ture different serial dependencies in the way symbols are distributed. Simple RNNs are trained with the same delta rule used for simple percep- trons. They learn to make context-dependent predictions that approximate the conditional probabilities of ensuing elements. To illustrate, the conditional probability of having ‘b’ immediately following ‘a’ (or pðbjaÞ) is estimated by paðÞ, b =paðÞ. Since the connection between an input node representing ‘a’ and an output node representing ‘b’ is strengthened by equation 2 in proportion to how often the bigram ab is found in input (and the learning rate γ), the net- work’s prediction of ‘b’ given ‘a’ will be a direct function of paðÞ, b . However, due to equation (2), the network learns nothing about – say – pcðÞ, b or paðÞ, c , from the distribution of ab’s. The assumption is in line with a purely associative view of error-driven learning and is supported by the intuitive observation that serial predictions can rely on already processed symbols only. Discriminative learning couples equation (2) with the idea that multiple cues are constantly in competition for their predictive value for a given outcome. Accordingly, learning proceeds not by simply associating co-occurring cues and outcomes, but by discriminating between multiple cues. In Rescorla-Wagner’s (1972) equations, the associative step is complemented with a competition-driven step, which forces the associative strength to take a step down when the outcome is present and the cue is not. Recent applications of Rescorla-Wagner equations to modelling a number of language tasks (see Pirrelli et al. 2020, this volume) are based on estimating the connection weights between two levels of representa- tion: one for raw strings, based on n-gram encoding, the other one for the sym- bolic encoding of morpho-lexical and morphosyntactic units. 262 Claudia Marzi et al.
Temporal Self-Organizing Maps (TSOMs) have recently been proposed to memorize symbolic time-series as chains of specialized processing nodes, that selectively fire when specific symbols are input in specific temporal contexts (Ferro et al. 2011; Marzi et al. 2014; Pirrelli et al. 2015). TSOMs consist of: 1. one layer of input nodes (where an input stimulus is encoded), 2. a bidimensional grid of processing nodes (the map proper), 3. two levels of independently trainable connections: a. input connections b. temporal connections.
Through the level of input connections, information flows from the input layer to all map nodes (one-way input connections). A second level of connectivity goes from each map node to any other node on the map (including itself), form- ing a pool of re-entrant temporal connections that update each map node with the state of activation of the map at the previous time tick (one-time delay). Like with RNNs, a word is input to a TSOM as a time series of symbols, one sym- bol at a time. At each time tick, activation spreads through both input and tem- poral connections to yield an overall state of node activation, or Map Activation
Pattern, at time t. In particular we use MAPtðÞs to refer to the Map Activation Pattern relative to the current stimulus s. The node with the top-most activation level in MAPtðÞs is referred to as Best Matching Unit, or BMUtðÞs . Finally, a time series of sequentially activated BMUs, namely
Steps (i) and (ii) incrementally enforce node specialization. The map tends to allocate maximally distinct nodes for the processing of (sub)strings, as a func- tion of their frequency in training. To illustrate the effects of node specialization vs. sharing on the representation of inflected forms of the same paradigm, Figure 2 sketches two possible end states in the allocation of BMU chains re- sponding to four German forms of beginnen ‘begin’: BEGINNEN (infinitive, 1p and 3p present indicative), BEGINNT (3s and 2p present indicative), BEGANNT (2p preterite) and BEGONNEN (past participle). In the left panel, BMUs are Inflection at the morphology-syntax interface 263
arranged in a word tree. At any node ni, one can always retrace backwards the nodes activated to arrive at ni. The right panel of Figure 2, on the other hand, offers a compressed representation for the three words, with letters shared by the three forms activating identical BMUs. As a result, when the shared node ‘N’ is activated, one loses information of which node was activated at the previ- ous time tick.
Figure 2: A word node tree (a) and a word node graph (b) representing German ‘#BEGINNEN$’, ‘BEGINNT$’, ‘BEGANNT$’ and ‘#BEGONNEN$’. Vertices are specialized nodes and arcs stand for weighted connections. ‘#’ and ‘$’ are, respectively, the start-of-word and the end-of-word symbol.
Let us briefly consider the implications of the two BMU structures for word processing. In the word-tree (left), BEGINNEN, BEGINNT, BEGANNT and BEGONNEN are perceived by the map as distinct forms at the earliest branch- ing point in the hierarchy (the ‘G’ node). From that point onwards, the four 264 Claudia Marzi et al. words activate three distinct node paths, which further bifurcate into a fourth path at the point where BEGINNEN and BEGINNT become distinct. Clearly, whenever a node has one outgoing connection only, the TSOM has no uncer- tainty about the ensuing step to take, and can anticipate the upcoming input symbol with certainty. In the word-graph on the right, on the other hand, branching paths converge to the same node as soon as the four input forms share an input symbol. Having more branches that converge to a common node increases the processing uncertainty by the map. The node keeps mem- ory of many preceding contexts, and its possible continuation paths are multi- plied accordingly. As we will see in the following section, this dynamic plays a key role in word acquisition and paradigm organization with TSOMs.
7.3 Processing inflection
7.3.1 The pace of acquisition
Token frequency is known to pace the acquisition of content words (i.e. exclud- ing function words such as articles, prepositions and conjunctions), particularly at early stages of language development (Huttenlocher et al. 1991; Goodman et al. 2008; Rowe 2012). It is widely accepted that speakers have fairly accurate knowledge of the relative frequencies with which individual verbs appear in dif- ferent tenses, or with different combinations of person and number features (Ellis 2002). Even if it is clear that speakers do not actively engage in consciously counting features, they nevertheless are very good at estimating frequency distri- butions and their central tendencies. Early acquisition of frequent words is generally understood to be a memory effect: the more frequently a word is input, the more deeply entrenched its stor- age trace in the speaker’s mental lexicon, and the quicker its access. In TSOMs, word processing is intimately tuned to word storage. Nodes that are repeatedly activated by an input string, become increasingly specialized for processing that string. At the same time, they are the memory nodes used for its long-term representation. The reason why frequent words are acquired at earlier stages can be better understood in terms of this processing-storage dynamic. To investigate how word frequency distributions affect the acquisition of in- flected words in a discriminative recurrent neural network, Marzi and colleagues (Marzi et al. 2014, 2016, 2018) ran a series of experiments where the inflectional systems of various languages (English, German, Italian, Modern Greek, Spanish, Standard Modern Arabic) are acquired by a TSOM trained on different frequency distributions of the same data. In the experiments, the acquisition of each input Inflection at the morphology-syntax interface 265 word form is timed by the epoch when the word is recalled correctly from its memory trace on the map.7 As an example, Figure 3 shows the pace of acquisition of verb forms from three German paradigms: a regular one (brauchen ‘need’), a moderately irregu- lar one (wollen ‘want’), and a highly irregular one (sein ‘be’). In the plots, circles represent word forms, ranked on the y-axis by increasing frequency values, and plotted on the x-axis by their epoch of acquisition. Each panel shows the out- come of two training regimes: white circles correspond to forms that are pre- sented to a TSOM 5 times each (uniform training regime); black circles correspond to the same forms when they are presented to a TSOM with their (scaled) frequency distributions in a reference corpus (realistic training regime). Acquisition times are averaged over 5 repetitions of the same experimental con- dition in the two training regimes.
braucht (18) will (84) ist (1001) brauchen (15) wollen (72) war (440) sind (400) brauchten (4) wollte (56) waren (116) brauchte (4) wollten (16) sein (65) brauche (4) willst (6) bin (49) gebraucht (3) wollt (3) gewesen (48) brauchst (2) wolltest (2) bist (13) seid(4) brauchtet (1) gewollt (2) warst (3) brauchtest (1) wolltet (1) wart (1) brauchend (1) wollend (1) seiend (1) 10 15 20 10 15 20 10 15 20 learning epoch learning epoch learning epoch
Figure 3: TSOMs’ learning epochs for brauchen ‘need’, wollen ‘want’, sein ‘be’ with uniform (white circles) and realistic (dark circles) training conditions. For each form, frequencies are given in brackets.
We observe a significant interaction between the pace of word acquisition in the two training regimes and the degree of inflectional (ir)regularity of a para- digm. In a regular paradigm like brauchen (Figure 3, left panel) word forms are acquired by a TSOM at about the same epoch: between epoch 13 and 15 in a
7 Intuitively, a word memory trace in a TSOM is the “synchronous” union set of the map acti- vation patterns (MAPs) for all symbols making up the word, or Integrated Activation Pattern (IAP). For a sequence of symbols to be recalled accurately from a memory trace, each BMU must contain detailed information about each symbol and its position in the target word. If either information is wrong, the sequence is wrongly recalled. 266 Claudia Marzi et al. uniform training regime (white circles), and between epoch 14 and 17 in a real- istic training regime (black circles). Note that, on average, words are learned earlier when they are uniformly distributed. The latter effect is perceivably re- duced in wollen (Figure 3, center panel), and reversed in sein (Figure 3, right panel), where realistically distributed items (black circles) tend to be learned more quickly than when the same items are presented five times each (white circles). Furthermore, the time span between the first and the last acquired form in the same paradigm is fairly short in brauchen, longer in wollen, and even longer in sein. Prima facie, these results are compatible with a dual-route account of regular and irregular inflection (e.g. Pinker and Ullman 2002). One could argue that the short time span taken to acquire regular forms supports the nearly instantaneous application of a general rule to the paradigm, following the acquisition of its stem. Likewise, the prolonged time span taken for the acquisition of an irregular paradigm is evidence that irregular forms are memorized in a piecemeal, item- ized fashion. However, since no rule learning is in place in a TSOM, a different generalization mechanism must be invoked to account for this evidence. In a TSOM, neighboring words (e.g. walking and walked,orwalking and speak- ing), trigger partially overlapping memory traces, i.e. integrated activation patterns that share a few nodes. Entrenchment of shared nodes benefits from cumulative exposure to redundant input patterns, making a TSOM sensitive to sublexical structures in the input. Conversely, non-shared nodes in partially overlapping memory traces compete for synchronous activation primacy in processing, and play an important role in extending inflection analogically across paradigms. To understand how inter-paradigmatic analogical extension takes place, sup- pose that the (regularly) inflected form walking is presented to a TSOM for the first time. Its overall integrated activation pattern takes advantage of full activation of the stem walk-from– say – the already known walks,andtheactivationofing- nodes trained on other forms like speaking or making.Notethating-nodes will compete with the s-node of walks, prospectively activated by the forward temporal connections emanating from walk-nodes. A successful generalization ultimately depends on the final outcome of this competition, based on the synchronous acti- vation of map nodes by the two different layers of TSOM connectivity: the input layer and the temporal layer. The comparatively faster process of acquisition of the inflected forms in a regular paradigm takes place when all inflectional endings are repeatedly seen across several paradigms,8 and the paradigm-specific stem is
8 We discuss processing effects of the frequency distribution of forms sharing the same inflec- tional ending (or inflectional entropy) in Section 7.3.2. Inflection at the morphology-syntax interface 267 already acquired. This generalization step is markedly more difficult within irregu- lar paradigms, where more stem allomorphs are available and compete with one another for activation primacy (e.g. sings, sang, sung). This is a gradient irregular- ity effect: the more stem allomorphs are found in a paradigm, the longer the time needed to acquire and associate them all with their endings. Note finally the important difference in the pace of acquisition between the two distributions of sein, in sharp contrast with the strong correlation between the two distributions of brauchen. Despite the difference in the order of magni- tude between the two distributions, the comparative insensitivity of regulars to frequency effects is reminiscent of a regularity-by-frequency interaction (Ellis and Schmidt 1998). In inflection, being more regular means being repeatedly at- tested in large classes of verb paradigms, and, within each paradigm, across all paradigm cells. This is not the case for irregular inflection. A radically suppletive paradigm like sein makes it hardly possible to infer an unattested form from other members of the same paradigm. The majority of sein forms are acquired one by one, as an inverse function of their own frequency distribution (and length): high-frequency items are learned earlier (Figure 3, black circles in the top left corner of the rightmost panel), and low-frequency items tend to be learned later. Conversely, in a paradigm like brauchen, the cumulative token fre- quency of the invariant stem brauch- compensates for the varying rate at which individual forms are shown in input. This is also true of affix allomorphs. The frequency of regular affixation is multiplicatively affected by an increase in the number of lexemes in training. A doubling in the number of verb entries in train- ing results (if we ignore phonological adjustments) in doubling the number of affixes they select. With irregular verbs, this is hardly the case, as affix allo- morphs thinly spread across irregulars, clustering in small subclasses, and incre- menting their frequency rather unevenly. Thus, the pace of acquisition of regular paradigms will be less sensitive to token frequency effects of single forms, since it can benefit from the cumulative boost in frequency of other forms in the same paradigm (for stems) and other forms of other paradigms (for affixes). All these competition effects point to the important role that paradigm en- tropy plays in word processing/learning. We already mentioned paradigmatic entropy in connection with child acquisition of inflection (Section 6.2).
7.3.2 Inflectional regularity: A processing-oriented notion
How degrees of inflectional regularity affect processing strategies has been recently evaluated cross-linguistically by looking at the way a TSOM predicts an upcoming input form in word recognition (Marzi et al. 2018; Marzi, Ferro and Pirrelli 2019). 268 Claudia Marzi et al.
The panels in Figure 4 illustrate the dynamic of word access for verb sets of 6 lan- guages: Standard Modern Arabic, English, German, Modern Greek, Italian and Spanish, which offer evidence of graded levels of morphological (ir)regularities and complexity. For each language, they plot how easily regularly vs. irregularly inflected verb forms are predicted by a TSOM trained on 50 high-frequency sub- paradigms, each containing 15 uniformly distributed forms. Symbol prediction is graphed by the distance of each symbol tothemorphemeboundarybetweenthe word stem and its inflectional ending (centered on the first symbol of the ending: x = 0). Each plot shows, on the y-axis, how well a symbol is predicted by the map on the basis of its preceding context. Symbol prediction in regular forms is plotted by green dashed lines, and symbol prediction in irregular forms by red solid lines. In all tested languages, the stem prediction rate increases more steadily in regulars than irregulars (negative values on the x-axis, p-value <.001). The trend reflects the reduction in uncertainty that the map experiences when serial infor- mation is received (like in spoken word recognition). As the signal unfolds, the set of possible competing sequences that are compatible with the signal narrows down, until the point is reached when only one candidate can match the incom- ing signal (so-called “uniqueness point” in Marslen-Wilson’s cohort model, 1990). For sure, not all stem allomorphs of a paradigm are equally likely to be compatible with an incoming signal at any given point in time. Competition among available candidates is modulated by frequency, with more frequent allomorphs being more entrenched and so more likely to win over their compet- itors. The result is that the map is slower to recognize low frequency allomorphs with high-frequency competitors, as the latter are harder to eliminate (Lively, Pisoni and Goldinger 1994; Luce 1986; Luce and Pisoni 1998). Turning back to Figure 4, another significant cross-linguistic effect is the drop in the prediction rate at the morpheme boundaries (x = 0), in both regularly and irregularly inflected forms. We take such a level of discontinuity at the mor- pheme boundary to reflect the map’s sensitivity to syntagmatic word structure. Harris’ Mathematical Structures of language (1968) describes how sublexical structures can be segmented out of a continuous string of symbols using the sequential likelihood between adjacent symbols. This work provided the founda- tions for much of later information-theoretic work on the bootstrapping of lin- guistic knowledge from unsupervised data (Brent 1999; Christiansen et al. 1998; Juola 1998). Our evidence is in keeping with this work. Since our training data contain no information about morphological structure, word structure emerges as an effect of specialization of processing nodes through learning. By being repeatedly exposed to input sequences, BMUs develop a context-sensitive repre- sentation of input symbols through incoming temporal connections, while devel- oping at the same time strong expectations for symbols yet to come, through Inflection at the morphology-syntax interface 269
Arabic English 3 3
2 2
I I R R
PREDICTION 1 PREDICTION 1
0 0 –6 –3 03 –6 –3 03 distance to morpheme boundary distance to morpheme boundary German Greek 3 3
2 2
I I R R
PREDICTION 1 PREDICTION 1
0 0 –5 03 –5 0 distance to morpheme boundary distance to morpheme boundary
Italian Spanish 3 3
2 2
I I R R
PREDICTION 1 PREDICTION 1
0 0 –8 –4 0 4 –5 0 5 distance to morpheme boundary distance to morpheme boundary
Figure 4: For each language set, regression plots of interaction effects between morphological (ir)regularity and distance to morpheme boundary, in non-linear models (GAMs) fitting the number of symbols predicted by TSOMs. Categorical fixed effect is regularity (green dashed lines) vs. irregularity (red solid lines). 270 Claudia Marzi et al. outgoing temporal connections. In the end, discontinuity in the strength of local temporal connections correlate with (graded) levels of sublexical structure. The drop of prediction rates at morpheme boundaries tends to be more promi- nent for regular stems than for irregular ones. This is a clear paradigm effect. Stem allomorphs are found in a few paradigm cells only, and select a subset of the in- flectional endings available in their paradigm. This reduces the amount of uncer- tainty at the morpheme boundary, as shown by the difference in prediction drop between regulars and irregulars in Figure 4. In particular, in Arabic irregular para- digms, inflectional endings are strongly predicted as a consequence of being cued by inflecting prefixes. Interestingly, from a cross-linguistic perspective, it can be observed that discontinuous patterns, typically attested in irregular paradigms of concatenative languages and systematically attested in non-concatenative mor- phologies such as Arabic, tend to require a higher processing cost of stems, and a lower cost in processing inflectional endings (Hahn and Bailey 2005). These results provide evidence that perception of morphological structure crucially interacts with formal transparency and regularity in all languages. As a general trend, sublexical constituents are perceptually more salient when they remain unchanged across different contexts. In addition, perception of structural discontinuity increases with the number of different contexts where constituents are found. In regular paradigms, stems and endings combine more freely than stems and endings in irregular paradigms do. Hence regulars tend to exhibit a clearer morphological structure than irregulars, which, in turn, tend to induce a more holistic processing strategy. As observed with stems, also inflectional endings tend to be predicted bet- ter by a TSOM as more symbols are input. Once more, this is mainly an effect of increasingly reduced processing uncertainty due to the narrowing down of the set of possible inflectional endings compatible with the input sequence. In the end, a uniqueness point is reached, i.e. a processing point where only one can- didate inflectional ending is compatible with the unfolding signal, and the whole input form is recognized. Balling and Baayen (2008, 2012) call the point at which an inflected form can be told from all other forms of the same para- digm Complex Uniqueness Point (CUP for short). They show that late CUPs elicit longer processing responses by human subjects than early CUPs do. As a result, we expect steeper prediction slopes for endings that are uniquely identified ear- lier, and less steep prediction slopes for endings that eliminate their potential competitors at a later stage. What we observe by looking at our data is that, most often, the facilitative effect of regularity on stem processing is partially reversed with inflectional endings. Endings that are selected by irregular stems tend to be predicted more easily than endings of regular stems. In fact, irregular stems are more discriminative for the class of endings they can possibly select, Inflection at the morphology-syntax interface 271 and this speeds up processing of ensuing endings, since they exhibit earlier CUPs than regulars do.
8 Measuring inflectional complexity
It makes a lot of intuitive sense to claim that some languages are inflectionally more complex than others. Everybody would agree that the English conjugation system is simpler than the German system, which is, in turn, simpler than the verb system of Modern Standard Arabic. However, when we try to motivate these deceptively trivial judgements, we are faced with a number of difficulties. Descriptive linguists have often approached the issue through comprehensive catalogues of the morphological markers and patterns attested in a given language (Bickel and Nichols 2005; McWorther 2001; Shosted 2006). According to such ap- proaches, the complexity of an inflectional system is assessed by enumerating the category values instantiated in the system and the range of available markers for their realization. The utility of such “enumerative” complexity or E-complexity (Ackerman and Malouf 2013) is however dubious on many counts. As already shown in Section 6, researchers from diverse theoretical perspec- tives observe that rich inflection in fact facilitates early morphological production. In competition-based (Bates and MacWhinney 1987), as well as functional (Slobin 1982, 1985) and cue-response discriminative perspectives (Baayen et al. 2011), non- ambiguous morphological paradigms such as those of Italian conjugation are ar- gued to provide better syntactic cues to sentence interpretation, as compared, for example, to the impoverished inflectional system of English verb agreement. Biuniqueness form-meaning relationships make inflectional markers more trans- parent, more compositional and in the end easier to be acquired than the one-to- many mappings of morphological forms to syntactic features that are found in English, Swedish and Dutch (Phillips 1995, 1996; Dressler 2010). Some researchers (e.g. Blom and Wijnen 2006; Crago and Allen 2001; Legate and Yang 2007) have focused on the amount of finite verbs that children receive from the adult input, to observe that the high percentage of overtly inflected forms correlates with the early production of finite forms by children. In the framework of Natural Morphology, Dressler and colleagues (Bittner et al. 2003) claimed that a richer in- flection makes children more aware of morphological structure, so that they begin to develop intra-paradigmatic relations sooner than children prompted by simpler systems do (as confirmed by the quantitative results in Xanthos et al. 2011). Another argument emphasizes that the logical problem of acquiring an in- flection system consists in learning not just the full range of formal markers, 272 Claudia Marzi et al. but the set of implicative relations between fully-inflected forms, which allow novel forms to be deduced from known forms (or the cell-filling problem, Ackerman and Malouf 2013; Ackerman, Blevins, Malouf 2009). To illustrate, suppose we have two hypothetical inflection systems, each with two categories only (say, singular and plural) and three different endings for each category: A, B, C for singular, and D, E and F for plural. In one system, paradigms are found to present three possible pairs of endings only:
9 Prediction across input words is calculated by incrementally assigning each correctly predicted symbol a 1-point score, i.e. the prediction score of the preceding symbol incremented by 1. Otherwise, for unpredicted symbols the score is 0. Therefore, the longer the input word, the more likely it is to be predicted. In our set of data, Spanish verb forms tend to be longer than any others. Inflection at the morphology-syntax interface 275
whole words 3
2 Arabic English German Greek Italian PREDICTION 1 Spanish
0 –5 05 distance to morpheme boundary
stems inflectional endings 3 3
2 2
1 PREDICTION PREDICTION 1
0 0 –7.5 –5.0 –2.5 20 46 distance to morpheme boundary distance to morpheme boundary
Figure 5: Regression plot of interaction effects between languages and distance to morpheme boundary, in a GAM fitting the number of symbols predicted by TSOMs for input words (Top panel, adapted from Marzi et al. 2018), and for stems (left panel) and inflectional endings (right panel) separately. 276 Claudia Marzi et al. morpheme boundary. Nonetheless, language-specific processing effects are better understood by looking at the way inflection is formally realized in each language. In particular, German inflection presents fairly systematic processes of stem alternation, followed by a full set of embedded endings such as -e,-en,-end (e.g. beginn-e, beginn-en, beginn-end, respectively ‘(I) begin’, ‘(we, they) begin’, ‘begin- ning’ (present participle)). As a result, German stems and endings are predicted according to two reversed patterns, showing, respectively, a growing and a de- creasing prediction rate, as plotted in Figure 5 (left and right panels). The effect can be understood in terms of the distance between the word’s Uniqueness Point (UP, Marslen-Wilson 1984; Marslen-Wilson and Welsh 1978), i.e. the point in the input where there is only one possible lexical continuation of the currently acti- vated node chain, and the Complex Uniqueness Point, or CUP, i.e. the point where an inflected form can be distinguished from all its paradigm companions (see §7.22). The earlier the UP, the easier for a stem to be predicted. Likewise, the earlier the CUP, the better for an ending – and hence a whole input word – to be accessed. These disambiguation points have an inhibitory effect on both word ac- cess and prediction, as confirmed by evidence on reaction times in acoustic word recognition (Balling and Baayen 2008, 2012). UP disambiguates the input stem from other onset-overlapping stems (be they paradigmatically-related or not). CUP distinguishes a specific ending from any other possible candidate, i.e. it distin- guishes the input form from any other paradigmatically related form. Clearly, the fewer the endings that combine with a stem, the easier their processing. It is useful at this stage to focus on the interaction between processing com- plexity and inflectional regularity. In this connection, we propose investigating in- flectional complexity through a continuous, graded notion of paradigmatic (ir) regularity. For each target form, we consider its “stem-family size”, i.e. the number of paradigmatically-related forms that share, with our target, the same stem. In ad- dition, for each paradigm, we calculate its average stem-family size, defined as the average size of the stem families belonging to the paradigm. This average score provides a quantifiable, graded notion of “paradigm regularity” that can be used in place of the traditional, dichotomous classification of inflected forms as either regular or irregular. Figure 6 shows how the two measures affect symbol prediction by a TSOM in the serial processing of fully inflected forms (top panels), as well as their stems and inflectional endings separately (respectively, left and right panels). Note that there is a clear, facilitative effect of the family size on the processing of stems only: the greater the number of forms sharing the same stem, the easier their processing (i.e. the greater the number of predicted symbols). Conversely, when we consider full-forms and endings, there is an inhibitory effect of the family size: the more in- flected forms share the same stems, the greater the processing uncertainty due to Inflection at the morphology-syntax interface 277
full forms 2.01.51.0 2.01.51.00.50.0 PREDICTION PREDICTION 0.50.0 fitted values fitted values 0.2 0.8 1.4 05 15 paradigm regularity stem-family size
stems endings 2.01.51.0 2.01.51.00.50.0 2.01.51.0 2.01.51.0 PREDICTION PREDICTION PREDICTION PREDICTION 0.50.0 0.50.0 0.50.0 fitted values fitted values fitted values fitted values 0.2 0.8 1.4 05 15 0.2 0.8 1.4 05 15 paradigm regularity stem-family size paradigm regularity stem-family size
Figure 6: GAMs predicting for our set of languages (as categorical fixed effect) the number of symbols predicted by TSOMs: fixed effects are plotted separately as paradigm regularity and stem-family size for full-forms (top panels), stems (left panels), endings (right panels). In addition to these covariates, the three GAM models include as smooth effect the corresponding length: word length for the full-form model, stem and suffix length respectively for the stem and ending models. the larger set of possible inflectional endings compatible with the input sequence. Interestingly, a graded notion of paradigm regularity, ranging from idiosyncratic to regular paradigms, going through intermediate levels of (ir)regularity, positively correlates with our task. The upshot is that regularity favors entrenchment of stems, with an average facilitative effect on processing. Paradigm entrenchment, however, has a structural price to pay, with more complex (larger) inflection sys- tems being more difficult to process than simpler ones. In a functional perspective, the evidence offered here can be interpreted as the result of a balancing act between two potentially competing communicative requirements: (i) a recognition-driven tendency for a maximally contrastive sys- tem, where all inflected forms, both within and across paradigms, present the 278 Claudia Marzi et al. earliest possible uniqueness points (UP and CUP); and (ii) a learning-driven bias for a maximally generalizable inflection system, where, for each paradigm, all forms in the paradigm can be deduced from any one of its forms, or from the smallest possible set of known forms. Clearly, a maximally contrastive system would take the least effort to process, but would require full storage of all (un- predictable) items, thus turning out to be slow to learn. A maximally generaliz- able system, on the other hand, would be comparatively easier to learn, but rather inefficient to process, especially when it comes to low-frequency items. What we observe is that, although languages may vary in the way they distrib- ute processing costs across each single word due to the typological variety of the inflectional processes they resort to, if we measure the per-word processing cost as a linear function approximately interpolating prediction scores for the start-of-word and end-of-word symbols, we get values that are fairly similar. This observation is also compatible with another clear pattern shown by the data presented in this section. In each of our sample languages, the differ- ence between the processing cost of forms in regular paradigms and the proc- essing cost of forms in irregular paradigms shows a structure-sensitive profile. The higher processing cost of irregular stems is partially compensated by a lower cost in processing the inflectional endings selected by irregular stems. Once more, at the level of the whole word, these structural effects make the in- flectional system, from an information theoretic angle, as functional as possible to possibly contrasting processing requirements. It should nonetheless be appreciated that the facilitative effect of fully con- trastive paradigms is the result of the interaction of more factors, including com- plexity of the paradigm, word length and frequency distribution. Comparatively small differences have mostly been observed between languages that exhibit the same (or a comparable) number of morphosyntactic oppositions, which require the same (or a comparable) amount of syntactic contexts to be checked and inter- preted. When these conditions are not controlled, acquiring an inflection system with a larger number of contrasting forms may turn out to be harder than acquir- ing a simpler inflection system. For example, Basque verb agreement marks an inflected verb form with affixes for subject, direct object and indirect object case. The system is agglutinative, and the number of possible distinct affix combina- tions for ditransitive verbs soon gets very large (up to 102 different forms in the present indicative of the auxiliary). Quantitative evidence from child language in- flection shows that production of root infinitive is more frequent and prolonged in Basque than it is in a less inflectionally rich language like Spanish (Austin 2010, 2012). In fact, this may be a consequence of several concomitant factors. Basque paradigms have a much larger number of cells than Spanish paradigms have. Furthermore, the amount of syntactic context that must be processed for a Inflection at the morphology-syntax interface 279 child to check case assignment on the main verb form, is considerably larger than what is needed for Spanish. Once more, it is not one factor only, but the concomitant interaction of a number of factors that may be responsible for a more complex system, and account for slower inflection acquisition. Far from dealing with the full complexity of inflectional systems across lan- guages, we suggested here a departure from traditional approaches based on ei- ther the stocktaking of features and their markers, or a full grammatical description of an inflection system, which all seem to require a lot of knowledge about lexical/grammatical units as well as rules/processes for their recombina- tion/merging. Computer simulations of (discriminative) inflection learning offer a novel perspective on these issues, since they do not require that formal repre- sentations are already established. Our evidence naturally prompts the view that the overall complexity of an inflectional system is the resulting equilibrium state of a number of conflicting processing requirements and adaptive responses to task-dependent pressures (see also Marzi, Ferro and Pirrelli 2019).
9 Concluding remarks
Inflection is a fundamental area of word inquiry that lies at the interface be- tween morphology proper, i.e. knowledge of how words are shaped and inter- nally structured, and syntax, i.e. knowledge of what syntactic contexts make certain lexical shapes obligatory. The two dimensions are logically distinct and, in principle, independent. Nothing, in the specific way words are arranged syn- tagmatically, impinges on the way the same words are inflectionally marked. The great variety of formal means by which identical clusters of inflectional fea- tures are marked in morphology, both cross-linguistically and within the same language, bears witness to this autonomy, and lends itself reluctantly to being cast into combinatorial patterns of morpheme arrangement. In this chapter, we mainly focused on aspects of word realization, i.e. on knowledge of the way words are inflected, and on what factors influence speakers’ acquisition of this knowledge. Here, we recap a few take-home points. Of late, the time-honored idea that word forms are organized through paradig- matic families has proved to be extremely fruitful in accounting for important ef- fects of lexical organization and processing. Paradigms appear to organize word forms in a network of items in complementary distribution (or, as Saussure put it, in absentia). This network is controlled by two functionally-motivated, interacting principles. The first principle is discriminative: formal variantsmustbeableto mark the entire space of inflectional contrasts exhibited by a language. From this 280 Claudia Marzi et al. perspective, the more dissimilar an inflected form is from its paradigm compan- ions, the more effectively it is associated with its cluster of inflectional features (or paradigm cell). However, if paradigmatically related forms were all arbitrarily dif- ferent, a speaker would be in no position to interpret or produce novel forms. The second principle counterbalances this effect. It is implicational or predictive: pat- terns of variation tend to be interdependent in ways that allow speakers to predict novel forms on the basis of encountered forms. This is a hallmark of regular inflec- tion. Nevertheless, even the most suppletive or least predictable paradigms in a language typically present a few implicational patterns of formal redundancy. A discriminative/implicational account of the paradigm dimension sheds light on the graded nature of (ir)regularity and structure in inflectional systems. Morphological irregularity is not dysfunctional, but responds to a maximally con- trastive function in both word recognition and production. Since irregularly in- flected forms are typically isolated, and are acquired by being committed to memory, it is only to be expected that irregularity strongly correlates with token frequency. Regularly inflected forms, on the other hand, can benefit from re- peated patterns of intra-paradigmatic formal redundancy and are, therefore, also sensitive to family size (or type frequency) effects. Any inflectional system of average complexity typically presents a whole range of gradation along this continuum. Models that postulate a dichotomous classification of inflected forms between regulars and irregulars can only ac- count for somewhat ideal cases of particularly simple inflectional systems. Both morphological theory and computational morphology have laid considerable emphasis on graded patterns of inflectional generalization governed by lexical information. For decades, this has been one of the cornerstones of Lexicalist Morphology and has guided important computational work on inheritance lexi- cal networks such as DATR. The idea that both general and irregular inflec- tional patterns can be cast into formally uniform statements ultimately blurs the distinction between rules and lexical entries. In DATR, so-called lexical rules are expressed as statements containing free variables, which are bound to constant, local values within individual, idiosyncratic lexical entries. The idea of measuring the complexity of an inflectional system in terms of inferential uncertainty (or information entropy) represents an important recent development in paradigm-based approaches to inflection. For any given verb stem s, one can estimate how easily an unknown stem-affix combination s-aj can be predicted on the basis of an already encountered stem-affix combination s-ak for the same verb. In addition, we can also estimate how much reduction in uncertainty we get in guessing s-aj, when we know more stem-affix combina- tions for the same paradigm. Overall, formal irregularity is not randomly scat- tered across paradigm cells. An inflectional system tends to reduce the amount Inflection at the morphology-syntax interface 281 of uncertainty in mastering it. Besides, however complex and irregular, an in- flectional system tends to be organized in such a way that less predictable forms are usually more frequent than more predictable forms.10 It is such an implicatively organized system of patterns and subpatterns that effectively ad- dresses learnability issues, by constraining an otherwise unrestricted set of combinatorial options. Such a global functional property of inflectional systems is significantly missed by purely realizational models of inflection. A further step in the same direction is made by moving from inheritance net- works to recurrent neural networks (RNNs), where the paradigmatic organization of lexical forms is responsible for coactivation and competition of concurrently stored items in word recognition and production. The step has far reaching con- sequences on the way we look at word knowledge, as it shifts the research focus from what speakers know when they know inflection, to how speakers develop knowledge of inflection through input exposure. According to a learning-based perspective, redundant patterns are predominantly statistical, and even irregu- larities appear to be motivated by their frequency distribution in the system and the general-purpose learning strategies of the human language processor. All these issues are very different in character from the formal constraints on units, representations or rule systems proposed within theoretical and computational models. Nonetheless, they offer an insightful perspective on language architec- ture, and shed novel light on issues of inflectional complexity. The most influential legacy of connectionism for models of lexical processing is probably the idea that storage and processing are not segregated in functionally independent modules of the language architecture, but are better conceived of as two interlocked dynamics of the same underlying process. In processing an input stimulus, nodes respond with a short-term activation. Due to reinforcement and competitive specialization, specific nodes are trained to respond more and more strongly to a specific class of stimuli only, forming a long-term memory trace for that class. Nodes that are repeatedly fired at short time delays by the same time series of stimuli give rise to a long-term chain of nodes specialized for processing that series. To put it in terms of Hebb’s law of neural plasticity, nodes used to- gether wire together. Once more, nodes that get repeatedly activated in a
10 This is in fact connected with language usage and its functional relation to language change. One can hypothesise that high-frequency items are used more frequently and are thus more prone to being phonetically reduced (e.g. Bybee 2000; Jurafsky et al. 2002; Pluymaekers et al. 2005). Alternatively, it can be argued that high frequency be a consequence of the gram- maticalization of a lexical item, and its resulting light functional load (Hopper & Traugott 1993). Nevertheless, common currency in language usage can play a role in protecting high- frequency irregular forms from analogical levelling (Milizia 2015). 282 Claudia Marzi et al. processingroutineforaspecificinputword,arethesameunitsthatareusedfor the stored representation of that word. This relatively straightforward mechanism provides a causal link between input word frequency, degrees of entrenchment of lexical representations and processing effects. In line with neuro-anatomical evidence (Wilson 2001; D’Esposito 2007; Ma et al. 2014), RNNs model lexical working memory as the transient activation of long-term memory structures. From this perspective, word family relations are long term memory effects, based on concurrent storage of full forms. But there is more to it than just a memory effect. The two fundamental classes of redun- dant inflectional patterns, stems and affixes, give rise to different, interacting word families (namely, paradigms and inflectional classes), which appear to play an important role in the way individual forms are processed and perceived by the speakers. Focusing on paradigms first, when one member of a paradigm (say walks)is input to an RNN, other non-target memory chains such as walk, walked and walk- ing are synchronously activated, due to the shared nodes associated with the com- mon stem. The more regular the paradigm, the fewer its stem allomorphs, and the more entrenched, on average, their corresponding memory chains. Co-activation, however, raises uncertainty at the stem boundary, where many outgoing connec- tions project their expectations for different upcoming endings. Such a degree of uncertainty can be taken to be an information-theoretic correlate of morphological structure at the level of network connectivity. High entropy paradigms increase un- certainty at the stem boundary, and make their internal structure more salient. Conversely, low entropy families develop more holistic word chains. This explains why forms in regular paradigms are perceived more compositionally than irregular ones are. Unlike irregular paradigms, where each allomorph can appear in a sub- set of paradigm cells only, regular stems appear throughout their paradigms. Competition between paradigmatically unrelated forms, on the other hand, takes place with forms sharing the same inflectional ending (e.g. all ing-forms). In self-organizing RNNs, the strength of the connection linking – say – speak- and walk- to ing is controlled by the conditional probability of speak- and walk- given -ing. A high entropy distribution of ing-forms, corresponding to a more uni- form distribution, favors a more balanced allocation of weights on connections at the stem boundary. Conversely, high frequency ing-forms tend to proportion- ally strengthen their connections to -ing nodes, weakening the corresponding connections from their low frequency competitors. There is a clear relationship between high levels of uncertainty at the stem boundary and word processing effects. Other factors being equal, uni- formly distributed members of (high entropy) inflectional families will take equalorcomparabletimeforprocessing. Conversely, an unbalanced Inflection at the morphology-syntax interface 283 competition, with few family members occurring much more frequently than others, results in the RNN being slower to recognize low frequency forms with high frequency competitors, as the latter are harder to eliminate (Lively et al. 1994; Luce 1986; Luce and Pisoni 1998). This explains why higher entropy families are processed more easily (Baayen et al. 2011; Bertram et al. 2000; Moscoso et al. 2004; Kuperman et al. 2010). In realistic input conditions, inflectional endings are not distributed uni- formly (Blevins et al. 2017). Hence, for any paradigm, the most balanced distri- bution of its members is the one where inflected forms with high frequency endings are seen more often than inflected forms with low frequency endings. In fact, this distribution strengthens the inflected forms that appear with com- petitive (high frequency) endings, while weakening those that select weaker endings, and this explains the role of inflectional entropy on processing (Milin et al. 2009a, 2009b). Complex processing dynamics offer a novel perspective on assessment of complexity issues in typologically diverse inflection systems. Nowadays, com- puter simulations and non-linear models of data regression offer the opportu- nity to visualize time-bound effects of structure complexity on word processing at a considerable level of detail. One can thus inspect the non-linear trend of the processing cost of forms in both regular and irregular paradigms, as well as differential processing patterns in typologically different languages. Although languages may considerably vary in the way processing costs are apportioned within specific inflection systems, when processing costs are measured as a lin- ear function interpolating processing ease from the start of the word to its end, linear slopes for different inflection systems are fairly comparable, suggesting that inflection systems are, ultimately, the result of a balancing act among a number of potentially conflicting requirements and adaptive responses to task- dependent pressures. However logically independent, the paradigmatic and syntagmatic dimen- sions of inflection must functionally interact during language acquisition. A distributional, discriminative approach to learning appears to conflate the issues of how inflection is marked, and in what contexts marking applies. We already mentioned evidence that children learn words in chunks (MacWhinney 1976; Wilson 2003). Upon hearing contexts where the same verb is found in different morphosyntactic contexts (as in “SHE walkS” and “THEY walk”), the child is in a position to use information of the pronominal subject and the verb suffix to dis- criminate third person singular contexts from non-third person singular contexts, thereby discovering the relationship between S-inflection and SHE in pre-verbal position. Paradigmatic word relations are ultimately associated with contrastive points along the syntagmatic dimension. Likewise, paradigm cells can be viewed 284 Claudia Marzi et al. as grammatical abstractions over this multi-dimensional space of systematic distributional contrasts in context. Clearly, this requires that longer stretches of words be concurrently committed to memory and organized through superposi- tional memory chains. There is mounting evidence that this must be true. The human brain is un- derstood to “contain” not only morphologically simple words, but also inflected and derived forms, compounds, light verb constructions, collocations, idioms, proverbs, social routine clichés and pre-compiled routinized chunks maximizing processing opportunities (Jackendoff 2002). Recognition of idiomatic expres- sions and multi-word units provides strong evidence of a processing system that uses all available pieces of information as soon as possible to constrain memory search and speed up processing of the most highly expected input (Grimm et al. 2017; McCauley and Christiansen 2017; Vespignani et al. 2010). Likewise, deficits in the working memory span (e.g. in children with SLD) explain difficulties in the acquisition of inflection, especially for large embedding contexts. The formal and structural similarity between periphrastic inflection and idiomatic expres- sions (Booij 2010; Bonami 2015) bears witness to the functional interaction be- tween concurrent, redundant storage of multi-word chunks and inflectional marking in language acquisition. We still know comparatively little about the way this is implemented in the brain. Nevertheless, recent empirical and experi- mental evidence suggests that the brain might make use of relatively unlimited long-term memory resources to compensate for the relatively limited capacity of working memory (Tremblay and Baayen 2010). By storing a number of fre- quently needed/used multi-word units as holistic chunks, the human processor can augment its capacity by filling working memory slots with word chunks rather than individual words. More recent neuro-functional models of working memory as a limited attentional resource distributed flexibly among all items to be maintained during processing (Ma et al. 2014) can also take advantage of pre- compiled long memory chunks. In fact, since the latter are retrieved through long-term temporal connections, working memory resources can be more effi- ciently used to maintain inter-chunk connections. We believe that dynamic memory models such as those suggested by Ma and colleagues, which are based on the functional integration between working memory and long-term memory resources, will shift the theoretical debate away from the traditional dichotomy between word-centred vs. syntactically-oriented accounts of inflection. Future research will likely focus on scale-free mechanisms for concurrent memorization of time-series of symbols of different length, as well as scale-dependent effects of their concurrent, hierarchical organization in the human brain. Inflection at the morphology-syntax interface 285
References
Ackerman, Farrell, James P. Blevins, & Robert Malouf. 2009. Parts and wholes: Implicative patterns in inflectional paradigms. In James P. Blevins & Juliette Blevins (eds.), Analogy in grammar: Form and acquisition,54–82. Oxford: Oxford University Press. Ackerman, Farrell & Robert Malouf. 2013. Morphological organization: The low conditional entropy conjecture. Language 89 (3). 429–464. Ackerman, Farrell & Gregory Stump. 2004. Paradigms and periphrastic expressions. In Louisa Sadler & Andrew Spencer (eds.), Projecting morphology,11–54. Stanford: CSLI Publications. Albright, Adam. 2002. Islands of reliability for regular morphology: Evidence from Italian. Language 78. 684–709. Altmann, Gerry T. M. & Yuki Kamide. 2007. The real-time mediation of visual attention by language and world knowledge: Linking anticipatory (and other) eye movements to linguistic processing. Journal of Memory and Language 57 (4). 502–518. Amha, Azeb. 2001. The Maale language. Dissertation University of Leiden. Aronoff, Mark. 1976. Word formation in generative grammar. Cambridge: the MIT Press. Austin, Jennifer. 2010. Rich inflection and the production of finite verbs in child language. Morphology 20 (1). 41–69. Austin, Jennifer. 2012. The case-agreement hierarchy in acquisition: Evidence from children learning Basque. Lingua 122 (3). 289–302. Baayen, R. Harald, Tom Dijkstra & Robert Schreuder. 1997. Singulars and plurals in Dutch: Evidence for a parallel dual-route model. Journal of Memory and Language 37 (1). 94–117. Baayen, R. Harald & Peter Hendrix. 2017. Two-layer networks, non-linear separation, and human learning. In Martijn Wieling, Martin Kroon, Gertjan van Noord (eds.), From Semantics to Dialectometry. Festschrift in honor of John Nerbonne,13–22. London, College Publications. Baayen, R. Harald, Peter Hendrix & Michael Ramscar. 2013. Sidestepping the combinatorial explosion: Towards a processing model based on discriminative learning. Language and Speech 56. 329–347. Baayen, R. Harald, Petar Milin, Dušica FilipovićĐurđević, Peter Hendrix, Marco Marelli. 2011. An amorphous model for morphological processing in visual comprehension based on naive discriminative learning. Psychological Review 118 (3). 438–481. Baayen, R. Harald, James M. McQueen, Tom Dijkstra & Robert Schreuder. 2003. Frequency effects in regular inflectional morphology: Revisiting Dutch plurals. In R. Harald Baayen & Robert Schreuder (eds.), Morphological structure in language processing, 355–390. Berlin: De Gruyter Mouton. Balling, Laura W. & R. Harald Baayen. 2008. Morphological effects in auditory word recognition: Evidence from Danish. Language and Cognitive processes 23 (7–8). 1159–1190. Balling, Laura W. & R. Harald Baayen. 2012. Probability and surprisal in auditory comprehension of morphologically complex words. Cognition 125 (1). 80–106. Bane, Max. 2008. Quantifying and measuring morphological complexity. In Proceedings of the 26th West Coast Conference on Formal Linguistics.69–76. Bates, Elizabeth & Brian MacWhinney. 1987. Competition, variation and language learning. In B. MacWhinney (ed.), Mechanisms of language acquisition, 157–163, Hillsdale, NJ: Lawrence Erlbaum Associates. 286 Claudia Marzi et al.
Beard, Robert. 1995. Lexeme-Morpheme Base Morphology: A General Theory of Inflection and Word Formation. Cambridge: SUNY Press. Bertram, Raymond, R. Harald Baayen & Robert Schreuder. 2000. Effects of family size for complex words. Journal of Memory and Language 42. 390–405 Berwick, Robert C. 1985. The acquisition of syntactic knowledge. Cambridge, MA: MIT Press. Bickel, Balthasar & Johanna Nichols. 2005. Inflectional synthesis of the verb. Martin Haspelmath, Matthew S. Dryer, David Gil & Bernard Comrie (eds.), The World Atlas of Language Structures,94–97. Oxford: Oxford University Press Bittner, Dagmar, Wolfgang U. Dressler & Marianne Kilani-Schoch (eds.). 2003. Development of verb inflection in first language acquisition: A cross-linguistic perspective. Berlin/New York: Mouton de Gruyter. Blevins, James P. 2003. Stems and paradigms. Language 79 (2). 737–767. Blevins, James P. 2006. Word-based morphology. Journal of Linguistics 42 (3). 531–573. Blevins, James P. 2016. Word and paradigm morphology. Oxford: Oxford University Press. Blevins, James P., Petar Milin & Michael Ramscar. 2017. The Zipfian paradigm cell filling problem. Ferenc Kiefer, James P. Blevins & Huba Bartos (eds.), Perspectives on morphological organization: data and analyses, 139–158. Leiden: Brill. Blom, Elma & Frank Wijnen. 2006. Development need not be embarrassing: The demise of the root infinitive and related changes in Dutch child language. Manuscript, University of Amsterdam/Utrecht University. Bloomfield, Leonard. 1933. Language. University of Chicago Press, Chicago. Bonami, Olivier. 2015. Periphrasis as collocation. Morphology 25, 63–110. Bonami, Olivier & Sarah Beniami. 2017. Joint predictiveness in inflectional paradigms. Word Structure 9 (2). 156–182. Bonami, Olivier & Gilles Boyé. 2014. De formes en thèmes. In Florence Villoing, Sarah Leroy & Sophie David (eds.), Foisonnements morphologiques. Etudes en hommage à Françoise Kerleroux,17–45. Presses Universitaires de Paris Ouest. Booij, Geert. 1993. Against split morphology. In Geert Booij & Jaap van Marle (eds.), Yearbook of Morphology 1993.27–49. Dordrecht: Kluwer. Booij, Geert. 1996. Inherent versus contextual inflection and the split morphology hypothesis. In Geert Booij & Jaap van Marle (eds.), Yearbook of Morphology 1995.1–16. Dordrecht: Kluwer. Booij, Geert. 2010. Construction morphology. Oxford: Oxford University Press. Börjars, K., Nigel Vincent & Carol Chapman. 1997. Paradigms, periphrases, and pronominal inflection: a feature-based account. In Geert Booij & Jaap van Marle (eds.), Yearbook of Morphology 1996, 155–180. Dordrecht: Kluwer. Brent, Michael R. 1999. Speech segmentation and word discovery: A computational perspective. Trends in Cognitive Sciences 3. 294–301. Briscoe, Josie, Dorothy V.M. Bishop & Courtenay F. Norbury. 2001. Phonological processing, language, and literacy: A comparison of children with mild-to-moderate sensorineural hearing loss and those with specific language impairment. Journal of Child Psychology and Psychiatry, and Allied Disciplines 42. 329–340. Brown, Roger. 1973. A first language: early stages. Cambridge, MA: Harvard University Press. Bruck, Maggie & Stephen J. Ceci. 1999. The suggestibility of children’s memory. Annual review of psychology 50 (1). 419–439. Bybee, Joan. 1988. Morphology as lexical organization. Theoretical morphology 119–141. Bybee, Joan. 1995. Regular Morphology and the Lexicon. Language and Cognitive Processes 10(5). 425–455. Inflection at the morphology-syntax interface 287
Bybee, Joan. 2000. Lexicalization of sound change and alternating environments. In Michael B. Broe & Janet M. Pierrehumbert (eds.), Laboratory Phonology V: Acquisition and the Lexicon, 250–268. Cambridge: Cambridge University Press. Caballero, Gabriela & Alice C. Harris. 2010. A Working Typology of Multiple Exponence. In Ferenc Kiefer, Mária Ladányi & Péter Siptár (eds.), Current issues in morphological theory: (ir)regularity, analogy and frequency. Selected papers from the 14th International Morphology Meeting, Budapest, May 13–16, 163–188. John Benjamins. Cazden, Courtney B. 1968. The acquisition of noun and verb inflections. Child development 39 (2). 433–448. Chouinard, Michelle M. & Eve Clark. 2003. Adult reformulations of child errors as negative evidence. Journal of Child Language 30. 637–669. Christiansen, Morten H., Joseph Allen & Mark S. Seidenberg. 1998. Learning to segment speech using multiple cues: A connectionist model. Language and Cognitive Processes 13. 221–268. Clark, Eve V. 1987. The principle of contrast: A constraint on language. In Brian MacWhinney (ed.), Mechanisms of language aquisition.1–33. Hillsdale, NJ: Lawrence Erlbaum Associates. Clark, Eve V. 1990. On the pragmatics of contrast. Journal of child language 17(2). 417–431. Coltheart, Max, Kathleen Rastle, Conrad Perry, Robyn Langdon & Johannes Ziegler. 2001. DRC: A Dual Route Cascaded model of visual word recognition and reading aloud. Psychological Review 108. 204–256. Corbett, Greville G. & Norman M. Fraser. 1993. Network Morphology: A DATR account of Russian nominal inflection. Journal of Linguistics 29. 113–142. Crago, Martha B. & Shanley Allen. 2001. Early finiteness in Inuktitut: The role of language structure and input. Language Acquisition 9 (1). 59–111. D’Esposito, Mark. 2007. From cognitive to neural models of working memory. Philosophical Transactions of the Royal Society of London B: Biological Sciences 362 (1481). 761–772. de Jong, Nivja H., Laurie B. Feldman, Robert Schreuder, Matthew Pastizzo & R. Harald Baayen. 2002. The processing and representation of Dutch and English compounds: Peripheral morphological and central orthographic effects. Brain and Language 81 (1). 555–567. De Pauw, Guy & Peter W. Wagacha. 2007. Bootstrapping morphological analysis of Gĩkũyũ using maximum entropy learning. Proceedings of the 8th Annual Conference of the International Speech Communication Association (INTERSPEECH 2007), Antwerp. 1517–1520. DeLong, Katherine A., Thomas P. Urbach & Marta Kutas. 2005. Probabilistic word pre- activation during language comprehension inferred from electrical brain activity. Nature Neuroscience 8 (8). 1117–1121. Dressler, Wolfgang U. 2005. Word-formation in natural morphology. In Pavol Štekauer & Rochelle Lieber (eds.), Handbook of word-formation, 267–284. Dordrecht: Springer Dressler, Wolfgang U. 2010. A typological approach to first language acquisition. In Michèle Kail & Maya Hickmann (eds.), Language acquisition across linguistic and cognitive systems, 109–124. Ellis, Nick C. 2002. Reflections on frequency effects in language processing. Studies in second language acquisition 24(2). 297–339. Ellis, Nick C., & Richard Schmidt. 1998. Rules or associations in the acquisition of morphology? The frequency by regularity interaction in human and PDP learning of morphosyntax. Language and cognitive processes 13 (2–3). 307–336. 288 Claudia Marzi et al.
Elman, Jeffrey L. 1990. Finding Structure in Time. Cognitive Science 14(2). 179–211. Evans, Roger & Gerald Gazdar. 1996. DATR: A language for lexical knowledge representation. Computational Linguistics 22 (2). 167–216. Farmer, Marion. 2000. Language and social cognition in children with specific language impairment. Journal of Child Psychology and Psychiatry, and Allied Disciplines 41. 627–636. Ferro, Marcello, Claudia Marzi & Vito Pirrelli. 2011. A self-organizing model of word storage and processing: implications for morphology learning. Lingue e Linguaggio X(2). 209–226. Finkel, Raphael & Gregory Stump. 2007. Principal parts and morphological typology. Morphology 17.39–75. Freudenthal, Daniel, Julian Pine & Fernand Gobet. 2010. Explaining quantitative variation in the rate of Optional Infinitive errors across languages: a comparison of MOSAIC and the Variational Learning Model. Journal of child language 37(3). 643–669. Geertzen, Jeroen, James P. Blevins & Petar Milin. 2016. The informativeness of linguistic unit boundaries. Italian Journal of Linguistics 28(2). 1–24. Gillis, Steven. 2003. A case study of the early acquisition of verbs in Dutch. In Dagmar Bittner, Wolfgang U. Dressler & Marianne Kilani-Schoch (eds.), Development of verb inflection in first language acquisition: A cross-linguistic perspective, 171–203. Berlin/New York: Mouton de Gruyter. Goldsmith, John. 2001. Unsupervised learning of the morphology of a natural language. Computational Linguistics 27(2). 153–198. Goldsmith, John. 2006. An algorithm for the unsupervised learning of morphology. Natural Language Engineering 12. 1–19. Goodman, Judith C., Philip L. Dale & Ping Li. 2008. Does frequency count? Parental input and the acquisition of vocabulary. Journal of Child Language 35. 515–531. Grimm, Robert, Giovanni Cassani, Steven Gillis & Walter Daelemans. 2017. Facilitatory effects of muli-word units in lexical processing and word learning: a computational investigation. frontiers in psychology 8. https://www.frontiersin.org/articles/10.3389/ fpsyg.2017.00555/full (accessed 10 August 2019). Haegeman, Liliane. 1995. Root infinitives, tense, and truncated structures. Language Acquisition 4. 205–255. Hahn, Ulrike & Todd M. Bailey. 2005. What makes words sound similar? Cognition 97. 227–267. Halle, Morris. 1973. Prolegomena to a theory of word formation. Linguistic Inquiry 4. 451–464. Halle, Morris & Alec Marantz. 1993. Distributed morphology and the pieces of inflection. In Hale, K. & SJ Keyser (eds.), The View from Building 20. Cambridge, Mass: MIT Press. Hamann, Cornelia & Kim Plunkett. 1998. Subjectless sentences in child Danish. Cognition 69 (1). 35–72. Hankamer, Jorge. 1989. Morphological parsing and the lexicon. In William Marslen-Wilson (ed.), Lexical representation and process, 392–408. The MIT Press. Harm, Michael W. & Mark S. Seidenberg. 1999. Phonology, reading acquisition and dyslexia: insights from connectionist models. Psychological Review 106(3). 491–528. Harris, Alice C. 2009. Exuberant exponence in Batsbi. Natural Language and Linguistic Theory 27. 267–303. Inflection at the morphology-syntax interface 289
Harris, Zellig S. 1942. Morpheme alternants in linguistic analysis. Language 18. 109–115. Reprinted in Joos (1957), 109–115. Harris, Zellig S. 1951. Methods in Structural Linguistics. University of Chicago Press, Chicago. Harris, Zellig S. 1968. Mathematical structures of language. New York: Wileyand Sons. Hockett, Charles F. 1947. Problems of morphemic analysis. Language 23. 321–343. Reprinted in Joos (1957), 229–242. Hockett, Charles F. 1967. The Yawelmani basic verb. Language 43. 208–222. Hockett, Charles F. 1987. refurbishing our foundations: elementary linguistics from an advanced point of view. John Benjamins, Amsterdam. Hopper, Paul & Elizabeth Traugott. 1993. Grammaticalization. Cambridge: Cambridge University Press [20032]. Huttenlocher, Janellen, Wendy Haight, Anthony Bryk, Michael Seltzer & Thomas Lyons, T. 1991. Early vocabulary growth: Relation to language input and gender. Developmental psychology 27(2), 236. Jackendoff, Ray. 1975. Morphological and semantic regularities in the lexicon. Language 51. 639–671. Jackendoff, Ray. 2002. Foundations of language: Brain, meaning, grammar, evolution. Oxford University Press; Oxford. Jordan, Michael. 1997. Serial order: a parallel distributed processing approach. Advances in Psychology 121. 471–495. Juola, Patrick. 1998. Measuring linguistic complexity: The morphological tier. Journal of Quantitative Linguistics 3(5). 206–213. Jurafsky, Daniel, Alan Bell & Cynthia Girand. 2002. The role of the lemma in form variation. Papers in laboratory phonology VII. 4–34. Jurafsky, David, Alan Bell, Michelle Gregory & William D. Raymond. 2001. Probabilistic relations between words: Evidence from reduction in lexical production. In Joan L. Bybee & Paul Hopper (eds.), Frequency and the Emergence of Linguistic Structure, 229–254. Amsterdam: Benjamins. Kail, Robert & Laurence L. Leonard. 1986. Word-finding abilities in children with specific language impairment. Monographs of the American Speech-Language-Hearing Association 25. Kemps, Rachèl J.J.K., Mirjam Ernestus, Robert Schreuder & R.Harald Baayen. 2005. Prosodic cues for morphological complexity: The case of Dutch plural nouns. Memory & Cognition 33 (3). 430–446. Keuleers, Emmanuel & Walter Daelemans. 2007. Memory-based learning models of inflectional morphology: A methodological case-study. Lingue e Linguaggio 6(2). 151–174. Kibort, Anna. 2010. Towards a typology of grammatical features. In Anna Kibort & Greville G. Corbett (eds.), Features. Perspectives on a key notion in linguistics,64–106. Oxford: Oxford University Press. Kilani-Schoch, Marianne, Ingrida Balciuniene, Katharina Korecky-Kröll, Sabine Laaha & Wolfgang U. Dressler. 2009. On the role of pragmatics in child-directed speech for the acquisition of verb morphology. Journal of Pragmatics 41(2). 219–239. Kolmogorov, Andrej N. 1965. Three approaches to the quantitative definition of information. Problems of information transmission 1(1). 1–7. Kostić, Aleksandar, Tanja Marković & Aleksandar Baucal. 2003. Inflectional morphology and word meaning: Orthogonal or co-implicative domains? In R. Harald Baayen & Robert 290 Claudia Marzi et al.
Schreuder (eds.), Morphological Structure in Language Processing,1–44. Mouton de Gruyter, Berlin. Krajewski, Grzegorz, Lieven, Elena V.M. & Anna L. Theakston. 2012. Productivity of a Polish child’s inflectional noun morphology: A naturalistic study. Morphology 22 (1). 9–34. Kuperman, Victor, Raymond Bertram & R.Harald Baayen. 2010. Processing trade-offs in the reading of Dutch derived words. Journal of Memory and Language 62 (2). 83–97. Legate, Julie A. & Charles Yang. 2007. Morphosyntactic learning and the development of tense. Language Acquisition 14 (3). 315–344. Leonard, Laurence. 2014. Children with Specific Language Impairment. The MIT Press. Leonard, Laurence & Patricia Deevy. 2011. Input distribution influences degree of auxiliary use by children with specific language impairment. Cognitive Linguistics 22. 247–273. Leonard, Laurence, Patricia Deevy, Patricia Fey & Shelley Bredin-Oja. 2013. Sentence comprehension in specific language impairment: A task designed to distinguish between cognitive capacity and syntactic complexity. Journal of Speech, Language, and Hearing Research 56. 577–589. Lieber, Rochelle. 1992. Deconstructing morphology. Chicago: Chicago University Press. Lively, Scott E., David B. Pisoni & Stephen D. Goldinger. 1994. Spoken word recognition: Research and theory. In Morton A. Gernsbacher (ed.), Handbook of psycholinguistics, 265–301. San Diego, CA: Academic Press. Luce, Paul A. 1986. A computational analysis of uniqueness points in auditory word recognition. Perception and Psychophysics 39. 155–158. Luce, Paul A. & David B. Pisoni. 1998. Recognizing spoken words: The neighborhood activation model. Ear and hearing 19 (1). 1–36. Ma, Wei J., Masud Husain & Paul M. Bays. 2014. Changing concepts of working memory. Nature Neuroscience 17 (3). 347–356. MacWhinney, Brian. 1976. Hungarian research on the acquisition of morphology and syntax. Journal of Child Language 3. 397–410. MacWhinney, Brian. 1995. The CHILDES Project: Tools for analyzing talk. Lawrence Erlbaum, Mahwah, NJ. Marangolo, Paola & Costanza Papagno. 2020. Neuroscientific protocols for exploring the mental lexicon: evidence from aphasia. In Vito Pirrelli, Ingo Plag & Wolfgang U. Dressler (eds.), Word knowledge and word usage: a cross-disciplinary guide to the mental lexicon, 124–163. De Gruyter. Marcus, Gary F., Steven Pinker, Michael Ullman, Michelle Hollander, T. John Rosen, Fei Xu & Harald Clahsen. 1992. Overregularization in language acquisition. Monographs of the Society for Research in Child Development 57(4, Serial No. 228). Marelli, Marco & Marco Baroni. 2015. Affixation in semantic space: Modeling morpheme meanings with compositional distributional semantics. Psychological Review 122 (3). 485–515. Marslen-Wilson, William D. 1984. Function and process in spoken word recognition: A tutorial overview. In Herman Bouma & Don G. Bouwhuis (eds.), Attention and performance X: Control of language processes, 125–150. Hillsdale: Erlbaum. Marslen-Wilson, William D. 1987. Functional parallelism in spoken word-recognition. Cognition 25 (1–2). 71–102. Marslen-Wilson, William D. & Alan Welsh. 1978. Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology 10. 29–63. Inflection at the morphology-syntax interface 291
Marzi, Claudia, Marcello Ferro, Franco Alberto Cardillo & Vito Pirrelli. 2016. Effects of frequency and regularity in an integrative model of word storage and processing. Italian Journal of Linguistics 28 (1). 79–114. Marzi, Claudia, Marcello Ferro, Ouafae Nahli, Patrizia Belik, Stavros Bompolas & Vito Pirrelli. 2018. Evaluating Inflectional Complexity Crosslinguistically: a Processing Perspective. In Proceedings of 11th LREC 2018, Miyazaki, Japan. Paper 745. Marzi, Claudia, Marcello Ferro & Vito Pirrelli. 2014. Morphological structure through lexical parsability. Lingue e Linguaggio XIII(2). 263–290. Marzi, Claudia, Marcello Ferro & Vito Pirrelli. 2019. A processing-oriented investigation of inflectional complexity. Frontiers in Communication 4. 48, 1–23. https://doi.org/10.3389/fcomm.2019.00048 Matthews, Peter H. 1972. Inflectional Morphology: a theoretical study based on aspects of latin verb conjugation. Cambridge University Press, Cambridge. Matthews, Peter H. 1991. Morphology: an introduction to the theory of word-structure. Second Edition. Cambridge University Press, Cambridge. McCauley, Stewart M. & Morten H. Christiansen. 2017. Computational investigations of multiword chunks in language learning. Topics in Cognitive Science 9 (3). 637–652. McClelland, James L. & David E. Rumelhart. 1981. An interactive activation model of context effects in letter perception: Part 1. An account of Basic Findings. Psychological Review 88. 375–407. McDonough, Joyce. 2000. On the bipartite model of the Athabaskan verb. In Theodor Fernald & Paul Platero (eds.), The Athabaskan Languages: Perspectives on a Native American Language Family, 139–166. Oxford: Oxford University Press. McNamee, Paul & James Mayfield. 2007. N-gram morphemes for retrieval. Working Notes for the CLEF 2007 Workshop, Budapest. McWorther, John. 2001. The world’s simplest grammars are creole grammars. Linguistic Typology (5). 125–166. Milin, Petar, Laurie B. Feldman, Michael Ramscar, Peter Hendrix & R. Harald Baayen. 2017. Discrimination in lexical decision. PloS one 12 (2). e0171935. Milin, Petar, Dušica FilipovićĐurđević & Fermín Moscoso del Prado Martín. 2009a. The simultaneous effects of inflectional paradigms and classes on lexical recognition: Evidence from Serbian. Journal of Memory and Language 60 (1). 50–64. Milin, Petar, Victor Kuperman, Aleksandar Kostić & R. Harald Baayen. 2009b. Words and paradigms bit by bit: An information theoretic approach to the processing of paradigmatic structure in inflection and derivation. In James P. Blevins & Juliette Blevins (eds.), Analogy in grammar: Form and acquisition, 214–252. Oxford University Press. Milizia, Paolo. 2015. Patterns of syncretism and paradigm complexity: The case of old and middle Indic declension. In Matthew Baerman, Dunstan Brown & Greville G. Corbett (eds.), Understanding and Measuring morphological complexity, 167–184. Oxford: Oxford University Press. Montgomery, James W. 2004. Sentence comprehension in children with specific language impairment: Effects of input rate and phonological working memory. International Journal of Language & Communication Disorders 39. 115–133. Montgomery, James W. & Laurence B. Leonard. 1998. Real-time inflectional processing by children with specific language impairment: Effects of phonetic substance. Journal of Speech, Language, and Hearing Research 41. 1432–1443. 292 Claudia Marzi et al.
Moscoso del Prado Martín, Fermín, Aleksandar Kostić & R. Harald Baayen. 2004. Putting the bits together: An information theoretical perspective on morphological processing. Cognition 94 (1). 1–18. Mulder, Kimberley, Ton Dijkstra, Robert Schreuder & R.Harald Baayen. 2014. Effects of primary and secondary morphological family size in monolingual and bilingual word processing. Journal of Memory and Language 72. 59–84. Noyer, Rolf. 1992. Features, positions and affixes in autonomous morphological structure. Ph.D. thesis, MIT. Paul, Hermann. 1920. Prinzipien der Sprachgeschichte. Max Niemayer Verlag, Tübingen. Perry, Conrad, Johannes C. Ziegler & Marco Zorzi. 2007. Nested incremental modeling in the development of computational theories: the CDP+ model of reading aloud. Psychological review 114(2). 273–315. Penke, Martina. 2012. The dual-mechanism debate. In Markus Werning, Wolfram Hinzen & Edouard Machery (eds.), The Oxford handbook of compositionality, 574–595. Oxford: Oxford University Press. Peters, Ann M. 1997. Language typology, prosody, and the acquisition of grammatical morphemes. The crosslinguistic study of language acquisition 5. 135–197. Phillips, Colin. 1995. Syntax at two: Cross-linguistic differences. In Carson Schuetze, Jennifer Ganger & Kevin Broihier (eds.), Papers on language processing and acquisition, 225–282. Cambridge, MA: MIT. Phillips, Colin 1996. Root infinitives are finite. In Andy Stringfellow, Dalia Cahana-Amitay, Elizabeth Hughes & Andrea Zukowski (eds.), Proceedings of the 20th annual Boston University conference on language development,588–599. Somerville, MA: Cascadilla Press. Pickering, Martin J. & Simon Garrod. 2007. Do people use language production to make predictions during comprehension? Trends in Cognitive Sciences 11 (3). 105–110. Pine, Julian M., Gina Conti-Ramsden, Kate Joseph, Elena V.M. Lieven & Ludovica Serratrice. 2008. Tense over time: Testing the Agreement/Tense Omission Model as an account of the pattern of tense-marking provision in early child English. Journal of Child Language 35. 55–75. Pinker, Steven. 1984. Language learnability and language development. Cambridge, MA: Harvard University Press. Pinker, Steven. 1989. Learnability and cognition. The acquisition of argument structure. Cambridge, MA: MIT Press. Pinker, Steven & Michael T. Ullman. 2002. The past and future of the past tense. Trends in cognitive sciences 6 (11). 456–463. Pirrelli, Vito. 2000. Paradigmi in morfologia. Un approccio interdisciplinare alla flessione verbale dell’italiano. Pisa-Roma: IEPI. Pirrelli, Vito & Marco Battista. 2000. The paradigmatic dimension of stem allomorphy in Italian verb inflection. Italian Journal of Linguistics 12 (2). 307–380. Pirrelli, Vito & Marco Battista. 2003. Syntagmatic and paradigmatic issues in computational morphology. Linguistica Computazionale XVIII–XIX. 679–702. Pirrelli, Vito, Marcello Ferro & Claudia Marzi. 2015. Computational complexity of abstractive morphology. In Matthew Baerman, Dunstan Brown & Greville G. Corbett (eds.), Understanding and measuring morphological complexity, 141–166. Oxford: Oxford University Press. Pirrelli, Vito, Claudia Marzi, Marcello Ferro, Franco Alberto Cardillo, Harald R. Baayen & Petar Milin. 2020. Psycho-computational modelling of the mental lexicon: a discriminative learning Inflection at the morphology-syntax interface 293
perspective. In Vito Pirrelli, Ingo Plag & Wolfgang U. Dressler (eds.), Word knowledge and word usage: a cross-disciplinary guide to the mental lexicon,21–80. De Gruyter. Pirrelli, Vito & François Yvon. 1999. The hidden dimension: a paradigmatic view of data-driven NLP. Journal of Experimental and Theoretical Artificial Intelligence 11. 391–408. Plaut, David C., James L. McClelland, Mark S. Seidenberg & Karalyn Patterson. 1996. Understanding normal and impaired word reading: Computational principles in quasi- regular domains. Psychological Review 103, 56–115. Plunkett, Kim & Patrick Juola. 1999. A connectionist model of English past tense and plural morphology. Cognitive Science 23 (4), 463–490. Pluymaekers, Mark, Mirjam Ernestus & R. Harald Baayen. 2005. Articulatory planning is continuous and sensitive to informational redundancy. Phonetica 62. 146–159. Popova, Gergana. 2010. Features in periphrastic constructions. In Anna Kibort & Greville G. Corbett (eds.), Features. Perspectives on a key notion in linguistics. Oxford: Oxford University Press, 166–184. Purdy, John D., Laurence B. Leonard, Christine Weber-Fox & Natalya Kaganovich. 2014. Decreased sensitivity to long-distance dependencies in children with a history of Specific Language Impairment: Electrophysiological evidence. Journal of Speech, Language, and Hearing Research 57 (3). 1040–1059. Ramscar, Michael & Melody Dye. 2011. Learning language from the input: Why innate constraints can’t explain noun compounding. Cognitive Psychology 62 (1). 1–40. Ramscar, Michael & Daniel Yarlett. 2007. Linguistic self-correction in the absence of feedback: a new approach to the logical problem of language acquisition. Cognitive Science 31 (6). 927–960. Ravid, Dorit, Emmanuel Keuleers and Wolfgang U. Dressler. 2020. Word storage and computation. In Vito Pirrelli, Ingo Plag & Wolfgang U. Dressler (eds.), Word knowledge and word usage: a cross-disciplinary guide to the mental lexicon, 582–622. De Gruyter. Rescorla, Robert A. & Allan R. Wagner. 1972. A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement. Classical conditioning II: Current research and theory 2. 64–99. Rice, Mabel L., Kenneth Wexler, Janet Marquis & Scott Hershberger. 2000. Acquisition of irregular past tense by children with specific language impairment. Journal of speech, language, and hearing research 43(5). 1126–1144. Rissanen, Jorma. 1989. Stochastic complexity in statistical inquiry. World Scientific Series in Computer Science. Singapore: World Scientific. Robins, Robert H. 1959. In defence of WP. Transactions of the Philological Society 58, 116–144. Reprinted in Transactions of the Philological Society 99. 116–144. Rowe, Meredith. 2012. A longitudinal investigation of the role of quantity and quality of child-directed speech in vocabulary development. Child Development 83. 1762–1774. Rumelhart, David & James McClelland. 1986. On learning the past tense of English verbs. In Rumelhart, D.E, McClelland J. & the PDP Research Group (eds.), Parallel distributed processing: Explorations in the microstructure of cognition.Vol.1.216–271, The MIT Press. Sadler, Louisa & Andrew Spencer. 2001. Syntax as an exponent of morphological features. In Geert Booij & Jaap van Marle (eds.), Yearbook of Morphology 2000,71–96. Dordrecht: Kluwer. Sagot, Benoît. 2018. Informatiser le lexique: Modélisation, développement et utilisation de lexiques morphologiques, syntaxiques et sémantiques. Habilitation, Université Paris- Sorbonne. 294 Claudia Marzi et al.
Sagot, Benoît, & Geraldine Walther. 2011. Non-canonical inflection: data, formalisation and complexity measures. In International Workshop on Systems and Frameworks for Computational Morphology.23–45. Springer, Berlin, Heidelberg. Scalise, Sergio. 1984. Generative morphology. Dordrecht: Foris. Schreuder, Robert & R. Harald Baayen. 1995. Modeling morphological processing. In Laurie B. Feldman (ed.), Morphological aspects of language processing, 131–156. Erlbaum. Shannon, Claude E. 1948. A mathematical theory of communication. Bell System Technical Journal 27. 379–423. Shosted, Ryan. 2006. Correlating complexity: a typological approach. Linguistic Typology (10). 1–40. Slobin, Dan I. 1982. Universal and particular in the acquisition of language. In Eric Wanner & Lila Gleitman (eds.), Language acquisition: The state of the art, 128–172. Cambridge: Cambridge University Press. Slobin, Dan I. (ed.). 1985. The crosslinguistic study of language acquisition. Hillsdale, NJ: Lawrence Erlbaum. Stark, Rachel & James Montgomery. 1995. Sentence processing in language-impaired children under conditions of filtering and time compression. Applied Psycholinguistics 16. 137–154. Stump, Gregory. 2001. Inflectional morphology: a theory of paradigm structure. Cambridge (UK): Cambridge University Press. Taatgen, Niels A. & John R. Anderson. 2002. Why do children learn to say “broke”? A model of learning the past tense without feedback. Cognition 86(2). 123–155. Tremblay, Antoine & R. Harald Baayen. 2010. Holistic processing of regular four-word sequences: A behavioral and erp study of the effects of structure, frequency, and probability on immediate free recall. In David Wood (ed.), Perspectives on formulaic language: Acquisition and communication, 151–173. Bloomsbury Publishing. Vespignani, Francesco, Paolo Canal, Nicola Molinaro, Sergio Fonda & Cristina Cacciari. 2010. Predictive Mechanisms in Idiom Comprehension. Journal of Cognitive Neuroscience 22 (8). 1682–1700. Wilson, Margaret. 2001. The case for sensorimotor coding in working memory. Psychonomic Bulletin & Review, 8 (1). 44–57. Wilson, Stephen. 2003. Lexically specific constructions in the acquisition of inflection in English. Journal of Child Language 30. 75–115. Wulfeck, Beverly & Elizabeth Bates. 1995. Grammatical sensitivity in children with language impairment. Technical Report CND-9512. San Diego: Center for Research in Language, University of California at San Diego. Wurzel, Wolfgang U. 1970. Studien zur deutschen Lautstruktur. Berlin: Akademie-Verlag. Xanthos, Aris, Sabine Laaha, Steven Gillis, Ursula Stephany, Ayhan Aksu-Koç, Anastasia Christofidou, Natalia Gagarina, Gordana Hrzica, F. Nihan Ketrez, F., Marianne Kilani- Schoch, Katharina Korecky-Kröll, Melita Kovacˇevic´, Klaus Laalo, Marijan Palmovic´, Barbara Pfeiler, Maria D. Voeikova & Wolfgang U. Dressler. 2011. On the role of morphological richness in the early development of noun and verb inflection. First Language 31(4), 461–479. Young, Robert & William Morgan. 1987. The Navajo Language. University of New Mexico, Albuquerque. Zipf, George K. 1935. The psychobiology of language. Houghton-Mifflin.